Basketball Trajectory Capture Method based on Neural Network Under the Background
of Sports Teaching
LiuHong
Copyright © The Institute of Electronics and Information Engineers(IEIE)
Keywords
Target detection, Trajectory prediction, Neural network
1. Introduction
In our daily life, there are all kinds of information around us all the time. As a
tool for human beings to capture, record and process all kinds of information, images
play an important role in human daily life. With the development of information technology
and computer technology, human beings are no longer satisfied with the simple application
of images [1]. They begin to try to use cameras and computers instead of human brains to detect
and track objects and process the captured images so that they can be used on more
occasions. Therefore, computer stereo vision has been formed. The research of computer
stereoscopic vision is the main research content in the field of computer vision,
which can be divided into binocular stereovision and multi binocular stereovision.
Among them, a multi binocular stereo vision system is a combination of binocular stereovision
systems [2]. In recent years, due to the continuous progress of science and technology, research
on computer stereo vision has become the focus. With the rapid development of the
modern social economy and the continuous improvement of material living standards,
people began to pursue a healthy lifestyle. Sports have gradually become a fashion.
Among many sports, basketball has become the most popular sport in China's sports
industry. Both the number of basketball courts and the number of basketball fans are
far ahead of other ball games.
Basketball has become a fashion. Basketball, as a group project, can improve people's
physical qualities, especially in the youth group. Basketball can help them develop
teamwork ability and optimize their physical qualities. Basketball is a competitive
sport, the game has lost and won. In order to win the game, we must improve the competitive
level of athletes [3]. The training of core strength can effectively improve the hit rate of jump shots,
and physical training can well prevent the damage caused by basketball. Additionally,
basketball footwork training is more significant but is frequently overlooked. The
NBA's use of small ball has led to the current game of basketball moving toward a
faster rhythm, a more adaptable style, and a more complete system. Basketball players
need to have outstanding feet and mobility technology because they must quickly switch
between attacking and defending [4]. Basketball footwork training is crucial because it helps players develop their physical
control and coordination, increase the stability of physical confrontation, successfully
finish each round's attack and defence, and develop a sound tactical system. In addition,
if we can get the running track of basketball players on the court and track it, we
can observe the coverage of players' running and the positions that often appear [5]. The tracking chart of players' running on the field can help the coaching team restore
the offensive and defensive paths of players on the field, and truly restore the implementation
of tactics. The positions of the defensive and offensive ends of the players shall
be arranged pertinently to give full play to the maximum value of each player. Basketball
footwork training and player track tracks complement each other, providing professional
theoretical analysis for basketball games. It can not only prevent players from injury
but also better design defense and attack strategies and improve the tactical system
[6]. The motion recognition system based on video images is shown in Fig. 1.
In the conventional sense, training consists of the coach watching the players train
or compete from the sidelines, recording it, assessing it, directing the players using
his or her years of teaching and training experience, and creating suitable training
schedules for them. The following drawbacks, however, are included in this analysis
and training approach [7]. First off, there are not enough coaches to match the number of players, which has
an impact on how effectively basketball players train. Second, it is challenging to
understand the underlying information that influences the players' footsteps when
we observe the players' training with human eyes because we cannot precisely determine
data such as the acceleration and angular velocity of the players' footsteps [8]. Finally, modern basketball is developing towards the trend of rapid movement, rapid
passing, and changing tactics. Coaches cannot track players all the time, so it is
inevitable that there will be some omissions. In a word, the traditional scheme has
been unable to adapt to the training and competition of modern basketball, and a more
intelligent, comprehensive solution is urgently needed [9]. This paper carries out the research on basketball target detection and rotating
ball trajectory prediction based on deep learning. By stacking the neural network,
we can achieve the task of predicting the trajectory, and meet the real-time and certain
accuracy.
In the current development context of AI technology, the combination of artificial
intelligence and robots is a trend, which also brings broader development space and
research value to robot research. At the same time, it also brings more research possibilities
to basketball robots. For the problems existing in the traditional methods mentioned
above, deep learning methods can all make good improvements. However, the detection
network of deep learning has a large computational load and requires a large amount
of training data, which brings some new problems. Therefore, this article aims to
address the shortcomings of traditional methods in basketball target detection and
trajectory prediction tasks. It attempts to combine deep learning methods with stronger
generalization and anti-interference abilities with the basketball visual system,
and takes into account the rotation characteristics of the basketball while predicting
the trajectory. Research on deep learning based basketball target detection and trajectory
prediction of rotating balls is conducted
The innovation of this article is to improve the accuracy of detection and ensure
sufficient detection speed on the basis of traditional methods, combined with intelligent
vision technology and motion capture technology, which has a certain supporting effect
on sports competition training
The paper's organization paragraph is as follows: The related work is presented in
Section 2. Section 3 analyzes the methods of the proposed work. Section 4, discusses
the experiments and results. Finally, in Section 5, the research work is concluded.
Fig. 1. The motion recognition system based on the video image.
2. Related Work
2.1 Research Status of Binocular Stereo Vision
With the development of computer science, optics, and image processing technology,
binocular stereo vision technology is constantly developing and will continue to be
practical and living. Early studies on binocular stereo vision were conducted in other
nations. As an illustration, the United States and Japan have led the globe in binocular
stereo vision research. China is still in the research and development stage at the
moment, necessitating the constant study, development, and improvement of a large
number of scientific researchers in order to advance stereo vision [10].
The research work on binocular stereo vision began in the 1960s. Roberts of Massachusetts
Institute of Technology built a building block world from multiple polyhedrons as
the detection environment, thus realizing the acquisition of three-dimensional information
from the two-dimensional environment, which means the birth of a stereo vision. The
Massachusetts Institute of Technology's Marr expanded the theoretical framework for
computer stereo vision in 1977 by proposing a technique for getting stereo pictures
from a parallax map. Since that time, computer vision has advanced [11]. Aspects like intelligent transportation, 3D range, robot navigation, virtual reality,
and real-time positioning and tracking are where binocular stereo vision is currently
most often used abroad. An intelligent traffic sensor was developed by MIT. The sensor
calculates the depth of field of the target through radar positioning technology and
binocular vision positioning technology and uses the method of image segmentation
to segment the position of the target [12]. This method solves the problem that the traditional image segmentation method cannot
be applied to the real-time high-speed movement of the target. Osaka University in
Japan has developed a binocular vision adaptive servo device, which uses the binocular
vision stereo ranging principle to take three relatively static objects in two images
as reference objects [13]. The movement direction of the target is predicted using the video image's Jacobian
determinant. The method achieves adaptive servo tracking of moving targets while dispensing
with the requirement of a standard servo tracking system that requires camera parameters.
The University of Tokyo has created a simulation robot dynamic path planning navigation
system by fusing real-time binocular vision with overall robot attitude data [14]. In 2005, the machine vision science laboratory of the University of Florida in the
central United States developed a cocoa system, which is based on the MATLAB platform
and can realize target detection and tracking under the dynamic background of the
UAV flight process [15]. The system first places the calibrated sensors that can be applied to computer stereo
vision in the experimental environment, then analyzes the image of one camera, the
difference between the images taken by two cameras and the coordinates of spatial
points, and finally realizes the calculation of the three-dimensional coordinates
of an object in space. The research status of binocular stereo vision is shown in
Fig. 2.
Although the development of binocular stereo vision in China is relatively late, many
research achievements have been made in recent years. Although great achievements
have been made in the research of binocular stereo vision technology in China, most
of the domestic research is on robot vision and industrial ranging, and the application
of binocular stereo vision in the field of sports competition is still relatively
lacking. Several high-speed cameras are built in the stadium for sports competitions.
Through the three-dimensional ranging system, the real-time tracking and three-dimensional
trajectory measurement of the target ball is realized [16]. This will be a great help to the training of athletes and the auxiliary judgment
of referees in current sports competitions such as table tennis.
Fig. 2. The research status of binocular stereo vision.
2.2 Current Situation of Basketball Game Video Analysis
In recent years, scholars at home and abroad have done some research work on basketball
game videos, mainly focusing on low-level visual feature extraction, scene classification,
video content analysis based on multi-feature fusion, specific semantic shot detection,
and so on [17]. There are numerous different analysis methods in use, however they can generally
be broken down into three categories: the analysis method is founded on the video's
underlying data. A technique for analysis based on the fusion of supplementary data
and fundamental video features. a method of multi-feature fusion-based semantic model
analysis.
Researchers use statistics of the main color distribution of the court in the basketball
game video image to segment the game video into the game and pause video clips, and
combine the shot duration to divide the video shots into game shots and nongame shots
[18]. The camera motion and basketball prior knowledge are extracted from the video for
high-level semantic analysis, and the fast break shot in the game is recognized from
the camera motion parameters in the basketball game shot. Scholars divide video into
wide-angle shots and close-up shots by extracting motion vectors from MPEG compressed
information [19]. For a wide-angle lens, further analyze the camera movement to mark specific video
content, such as steals, fast breaks, possible shots, etc. Through the video information
to distinguish the motion of the object, the motion of the basketball is tracked,
and the color feature and edge feature are used to automatically locate the backboard.
The Hough transform is used to detect the backboard and basket, and the relationship
between the basketball position and the backboard and basket is considered to judge
whether it is a dunk or a long-distance shot [20]. The motion information of the basketball object in the basketball game video is
extracted, and the semantic characteristics of its motion trajectory are analyzed.
2.3 Action Recognition Classification
In recent years, with the rapid development of computer vision technology and wearable
sensor technology, human daily activities, and action recognition have been widely
used in different scenes and fields, and become a hot research direction of many researchers
at home and abroad [21]. At this stage, many schemes have been used for motion recognition, of which the
two mainstream solutions are video image-based technology and wearable sensor-based
technology. As shown in Fig. 3, they will be introduced respectively below.
Generally, the recognition scheme based on video images is to segment the target object
from each frame of the image, extract the pose, action, and position of the target,
and then apply various classification algorithms for action recognition. The researchers
proposed a basketball event detection method based on multimodality [22]. By extracting audio and visual features from basketball videos, the feature mapping
was established with audio keywords, and the classifier recognized 9 basketball events.
The same is to detecting basketball events. Researchers use a group of filters to
respond to the motion pattern of each video frame, create an energy redistribution
function, establish a video analysis framework based on the Hidden Markov model, and
classify 16 basketball events [23]. The key point trajectory of the action is extracted into the divided space-time
sub-region, and the two-stage SVM is used to establish a high-dimensional training
framework [24]. This method has high recognition accuracy in detecting typical space-time actions
such as jump shots, standing shots, and layups. Different event recognition modules
are set to analyze the action and generate the status, position, and mode of each
object. In an environment with sufficient light and no occlusion, the scheme has a
good recognition effect. In addition, the framework based on deep learning is also
used in various motion recognition systems using video recording [25]. The researchers use the attention mechanism to continuously locate and track the
target players, then use a cyclic neural network to extract the action features, and
finally add an RNN to realize the action classification.
Reference [26] proposes a basketball trajectory tracking algorithm based on correlation filtering
and fusion of multiple features. This method extracts relevant features from the background
and target area of the basketball image, and achieves basketball trajectory tracking
through feature response maps. However, this method does not denoise the basketball
image and cannot accurately track the center coordinates of the basketball target,
resulting in low tracking accuracy; Reference [27] proposes a dribbling oriented shooting trajectory tracking method based on symmetric
algorithm. The maximum variance threshold method is used to partition the motion area
of basketball orientation, and the Camshift method is used to mark the trajectory
contour of dribbling orientation. Based on the contour extraction results, the grayscale
pixel value feature points of the video sequence are fitted, and a dribbling oriented
trajectory tracking model is established to accurately track the dribbling trajectory.
However, this method is affected by Gibbs artifacts, Unable to accurately track the
trajectory of basketball flight within the blind spot range
In order to obtain the best performance of motion recognition, multi-sensor and multi-algorithm
fusion have become a new development trend. The scheme based on wearable sensors collects
data through various sensors, processes and analyzes the data accordingly, and then
uses various classification algorithms to realize action recognition. Researchers
have proposed a wrist strap-based sensor system to identify the actions in basketball
games [28].
The traditional method of obtaining technical parameters is to add sensors to basketball
players, but the disadvantage of this method is that it may affect their game performance.
However, the recorded images of basketball games usually have a unified shooting mode,
and their fidelity and real-time interactivity provide strong support for obtaining
technical parameters of basketball players. This can enable basketball players and
coaches to achieve intuitive teaching and fast feedback, Furthermore, it can greatly
reduce the likelihood of athlete injuries. Therefore, various sports action recognition
and tracking technologies have been used to extract technical actions from these images,
achieving human-computer interaction and bringing tremendous results in further improving
athlete skills and protecting them from sports injuries
Fig. 3. Classification of action recognition methods.
3. Design of Application Model
In order to solve the problem of decreasing ability to detect small objects with increasing
network depth, a feature fusion network was constructed. The feature layers with rich
semantic information in the upper layer were fused with the feature maps with rich
object position information in the lower layer, so that the network depth increased
and the position information of the object could still be learned. Improved the network's
ability to detect small targets. The network in this article can adapt to different
lighting, environmental interference and other factors, with higher detection accuracy
than traditional object detection algorithms and detection speed that meets the requirements
of basketball vision systems.
3.1 Force Analysis and Motion Modeling
In this chapter, according to the existing research foundation, a motion model considering
rotation is proposed and the discrete form of the motion model is derived. Define
the world coordinate system. The Z-axis of the coordinate system is vertical to the
ground, the Y-axis is parallel to the long side of the table, the X-axis is parallel
to the short side of the ground, and the coordinate origin is the center point of
the court. Then gravity, air resistance, and Magnus force can be described as [29]:
When basketball flies in the air, the attenuation of rotation speed is very small,
so the rotation speed is regarded as a constant in this chapter. The magnitude of
air resistance is directly proportional to the square of the flight speed, and the
proportional coefficient is determined by the air resistance coefficient, air density,
and the cross-sectional area of the basketball. The direction of air drag is opposite
to the direction of flight speed. Given the comprehensive force, the motion model
of the rotating basketball can be derived from Newtonian mechanics as follows.
The continuous motion model can directly calculate the motion state of a basketball
at any time without iteration because it is aware of the initial motion state. This
effectively eliminates the iteration error and cut-off error and also directly describes
the relationship between the trajectory position time series and the initial motion
state. This provides a necessary basis for using the optimal mathematical method to
estimate the motion state of the rotating basketball. The formula is as follows [30].
In this paper, the Fourier series is used to fit the attenuation law of flight speed
with time. The formula is as follows.
The expression of the flight speed in the motion model relative to the initial motion
state can be solved [31].
According to the superposition of integrals, we can integrate the sub-functions in
the above formula respectively to obtain the following formula.
In order to ensure a high goal success rate, we must accurately predict the position
of the rotating basketball reaching the basket. We can deduce the constraint relationship
between the rotation speed and the flight speed of two adjacent frames as follows
[32].
The left part of the formula represents the acceleration caused by the Magnus force
calculated according to the motion model using the flight speeds of two adjacent frames.
The right part of the formula represents the constraint relationship between the current
flight speed and rotation speed in line with the current Magnus force definition.
The gnu’s force is perpendicular to the rotation speed, so we can also get:
It can be seen that the state estimation of rotation speed is essentially equivalent
to the state estimation of flight speed. The continuous motion model derived in this
paper essentially describes the motion law of basketball in a period of time under
the current motion state, which makes it possible to use the constraints of the motion
model to optimally estimate the motion state according to the observation information
of the continuous multi-frame trajectory position time series. Firstly, the curve
coincidence degree is defined as the sum of the Euclidean distance between the trajectory
prediction value and the trajectory observation value of consecutive multiple frames.
It can be seen that the smaller the sum of Euclidean distances, the higher the curve
coincidence between the trajectory prediction value and the trajectory observation
value. The sine and cosine functions, which are high-order nonlinear functions about
the motion state, are abundant in the motion model. The steps of dynamic state estimation
and trajectory prediction are shown in Fig. 4.
Fig. 4. The steps of dynamic state estimation and trajectory prediction.
3.2 Target Detection based on Feature Fusion Network
In basketball target detection tasks, a large amount of training data is also required
to train the network. Each image needs to be manually annotated, labeled, and labeled
with category and basketball position information. In the task of trajectory prediction,
basketball trajectory data is required. Creating these data requires a significant
amount of time and effort, and there is currently no publicly available large basketball
dataset available for use. Therefore, in order to complete the research in this article,
a large amount of basketball image data was manually collected under different environments,
colors, and lighting conditions, and further constructed the dataset required for
network training
The vision system's main responsibility with the basketball robot system is to quickly
and precisely recognise the target and position of the basketball. Mixup is the process
of combining two random images in a specific ratio. Additionally, the classification
will be distributed based on the size of each image. The implementation method is
as follows.
The implementation of Mixup is simple and does not increase the amount of computation.
It is one of the good data enhancement methods. In order to enhance the detection
and location ability of the network for small targets, it is necessary to fuse the
underlying feature information in the convolution down sampling process. The output
of each layer can be used to detect the object category and position. The FFN network
consists of two lines, a bottom-up line and a top-down line, which are horizontally
linked. This can make use of the underlying positioning details so that the network
can learn more accurate location information while learning the target features, especially
for the detection of small objects.
Structure diagram of feature fusion network is shown in Fig. 5.
Fig. 5. Structure diagram of feature fusion network.
The depth increase of the model also loses the position information of many objects
in the original image, which brings great difficulties to the detection of small objects.
It can be called the sensitivity of the basis of each neuron, which means that the
error will change as much as the basis changes, as shown in follow.
Compared with down sampling, up sampling is used to enlarge the resolution of the
image, and the quality of the enlarged image can exceed that of the original image.
In the structure of the traditional neural network, the model does not pay attention
to the impact of information processing at the previous time on the current time.
Therefore, this paper uses the LSTM network to predict the trajectory of basketball.
The partial code of this article is as follows:
frame nums # Total video frames
frame stride = mod (frame_nums ,16) # Take the step size, taking 16 frames as an example
frame_count =1 # Start Framecount = 1
while True:
if frame_count == count:
gray_frame # Grayscale processing of image frames
save frame # Save this frame
count += frame stride
frame count += 1
4. Results and Analysis
The training data in this paper are from 400 basketball tracks collected in the laboratory
and 1000 basketball tracks obtained on the network, of which 1200 are used as the
training data set and 200 as the test data set. In the aspect of prediction step size,
the network with step sizes of 5, 10, and 20 are trained with 1000 epochs, and the
experimental results are analyzed. The experimental error results are shown in Table 1.
As can be seen from the above table, when the prediction step size gradually increases,
the prediction error also increases. The possible reason is that in the training process,
the error will also be superimposed and will be enlarged with the increase of step
size, but the error range is within the acceptable range. In this paper, according
to different training steps, input the trajectory data of the first 15 frames, and
experiment with the error of the network when predicting 30, 40, 50, and 60 frames
to analyze the cumulative error of prediction with the increase of the number of frames.
The experimental error results are shown in Table 2 and Fig. 6 below.
The image set in this article mainly comes from regular competition video frames,
which are filtered and included in the database of this article
In terms of real-time performance, this paper calculates the time used to predict
the 40th and 60th frames under three steps to verify the feasibility of the network,
and compares it with the physical model method. It can be seen that the physical model
has great real-time advantage in prediction time; with the increase of step size,
the time required for network prediction is also increasing. The speed of forecasting
60 frames at a step size of 20 is 87.96 ms, which is considerably slower than the
network performance under the first two stages but still satisfies the real-time criteria.
Table 3 shows the time of the three steps in predicting 40 and 60 frames.
The purpose of trajectory prediction is to get an accurate hitting point which is
helpful to increase the success rate of returning the ball. In the prediction experiment
of hitting point, this paper compares the flight model and rebound model of basketball.
The predicted time-consuming results are shown in Fig. 7.
The three groups of straight-line trajectories in this experiment are 14 meters from
the bottom line position to the center line position of the basketball court. The
estimated distance of the three groups of straight-line trajectories is compared with
the real distance, and their errors are calculated. Basketball vision system needs
to accurately predict the trajectory of rotating basketball and judge the rotation
type of basketball. Although the traditional trajectory prediction method based on
physical model and motion modeling can meet the requirements of vision system, its
prediction accuracy, especially for the judgment of rotating ball, is not ideal. Experiments
show that the proposed network is closer to the real value in the accuracy of prediction
points, and the accuracy is greatly improved compared with the traditional algorithm.
The prediction speed meets the requirements of basketball vision system.
Compare and analyze the method proposed in this article with the method proposed in
reference [7], and calculate the accuracy of the two models in capturing basketball motion trajectories.
The experimental results are shown in Table 4 below.
The comparison of running time between the two models is shown in Table 5
From the above data statistics, it can be seen that the method model proposed in this
article has significant advantages compared to traditional models
Fig. 6. Error of three-step sizes in 30, 40, 50, and 60 frame prediction.
Fig. 7. The predicted time-consuming results.
Table 1. Error under three steps.
|
|
5
|
10
|
20
|
|
Training error
|
5.915
|
6.268
|
7.217
|
|
Test error
|
6.013
|
7.121
|
8.327
|
Table 2. Error of three-step sizes in 30, 40, 50, and 60 frame prediction.
|
Prediction frame
|
30
|
40
|
50
|
60
|
|
Step 5*10^3
|
7.30
|
8.40
|
9.70
|
11.20
|
|
Step 10*10^3
|
7.70
|
8.20
|
9.10
|
10.60
|
|
Step 20*10^3
|
8.20
|
9.10
|
10.20
|
10.80
|
|
Comparison method 1
|
11.30
|
15.60
|
21.30
|
28.90
|
|
Comparison method 2
|
12.50
|
17.30
|
24.70
|
31.50
|
Table 3. Time of the three steps in predicting 40 and 60 frames.
|
Prediction frame
|
40 frame time
|
60 frame time
|
|
Step size is 5 (MS)
|
21.30
|
44.80
|
|
Step size is 10 (MS)
|
33.60
|
69.30
|
|
Step size is 20 (MS)
|
45.20
|
87.90
|
|
Comparison method 1
|
17.60
|
35.30
|
|
Comparison method 1
|
16.30
|
33.80
|
Table 4. Statistics on the accuracy of basketball rotation trajectory recognition.
|
NO.
|
The method of this article
|
The method of reference [7]
|
|
1
|
83.96
|
84.87
|
|
2
|
89.84
|
83.15
|
|
3
|
90.18
|
83.32
|
|
4
|
88.29
|
79.80
|
|
5
|
87.22
|
80.03
|
|
6
|
88.88
|
85.45
|
|
7
|
87.35
|
82.79
|
|
8
|
83.17
|
84.96
|
|
9
|
90.83
|
78.68
|
|
10
|
85.86
|
80.39
|
Table 5. Comparison of Model Running Times.
|
Number of frames
|
The method of this article(ms)
|
The method of reference [7](ms)
|
|
100
|
102
|
211
|
|
500
|
350
|
521
|
|
1000
|
622
|
754
|
5. Conclusion
The vast expansion of sports video data has generated a great deal of interest in
content-based sports video analysis. Basketball game has emerged as one of the hottest
topics in the field of content-based video retrieval and analysis due to its significance
as a component of sports videos. An efficient description technique is designed to
extract video semantic elements in order to reduce storage requirements and meet the
individualized needs of consumers. In this topic, basketball video is used as the
analysis object. Basketball objects are found, tracked, and their motion tracks are
extracted. A specific region enhancement method based on the resolution feature map
is proposed. Firstly, the feature map of the image is extracted according to the selective
attention mechanism of the human eye to locate the object and then weighted with the
image gray to enhance the basketball candidate region. The basketball object template
is established according to the basketball detection results, and the two-level scaling
algorithm is used to search the search area globally in the subsequent frames. Furthermore,
combined with prior knowledge, the three-dimensional coordinates of basketball are
estimated, and the trajectory of basketball in three-dimensional space is reproduced.
The shortcomings and potential areas for this paper's improvement are listed below.
There will be more influence from the environment in the competition, and the original
richness of the data set in this work is still absent. Additionally, this paper's
network structure is appropriate for various ball sports as well as basketball situations.
It can be used for more ball game scenes as long as sufficient rich data sets are
created and retrained on the basis of this network. In the basketball recognition
system, the decision-making system and control system also play a vital role. Basketball
has an obvious relative motion to the camera, so it is inaccurate to reconstruct its
motion only from the single view image sequence. The next step is to collect video
data from multiple perspectives for reconstruction.
The proposed network has significant real-time advantages in predicting time, and
its accuracy in predicting points is closer to the true value. Compared with traditional
algorithms, the accuracy has been greatly improved. It can help basketball players
practice in daily training, saving training costs and improving training efficiency.
In subsequent basketball training, the system in this article can be used for auxiliary
training
In terms of tracking basketball trajectory, the target tracking algorithm proposed
in this article has improved tracking accuracy, but there are still certain errors.
Therefore, further research is needed on tracking algorithms for fast moving small
targets.
The algorithm architecture proposed in this article is aimed at basketball technical
actions. How to apply it to short video platforms, how to present it to the web, mini
programs, and how to recognize and label technical actions in videos and push them
to target users. These directions are also worth further research.
REFERENCES
Liu L, Hodgins J. Learning basketball dribbling skills using trajectory optimization
and deep reinforcement learning. ACM Transactions on Graphics (TOG), 2018, 37(4):
1-14.
Li B, Xu X. Application of artificial intelligence in basketball sport. Journal of
Education, Health and Sport, 2021, 11(7): 54-67.
Liu N, Liu P. Goaling recognition based on intelligent analysis of real-time basketball
image of Internet of Things. The Journal of Supercomputing, 2022, 78(1): 123-143.
Tsai W L, Pan T Y, Hu M C. Feasibility study on virtual reality based basketball tactic
training . IEEE Transactions on Visualization and Computer Graphics, 2020.
Wen P C, Cheng W C, Wang Y S, et al. Court reconstruction for camera calibration in
broadcast basketball videos . IEEE transactions on visualization and computer graphics,
2015, 22(5): 1517-1526.
Kang S Y, Park K H. Design and Implementation of Basketball Game Content based on
Unity Game Engine and Android Platform. The Journal of Digital Contents Society, 2020,
21(9): 1567-1573.
Min B J. Application of Monte Carlo simulations to improve basketball shooting strategy
. Journal of the Korean Physical Society, 2016, 69(7): 1139-1143.
Liu W, Yan C C, Liu J, et al. Deep learning based basketball video analysis for intelligent
arena application. Multimedia Tools and Applications, 2017, 76(23): 24983-25001.
Chen W, Lao T, Xia J, et al. Gameflow: narrative visualization of NBA basketball games
. IEEE Transactions on Multimedia, 2016, 18(11): 2247-2256.
Yang L, Wang B, Zhang R, et al. Analysis on location accuracy for the binocular stereo
vision system . IEEE Photonics Journal, 2017, 10(1): 1-16.
Jia Z, Yang J, Liu W, et al. Improved camera calibration method based on perpendicularity
compensation for binocular stereo vision measurement system. Optics express, 2015,
23(12): 15205-15223.
Luo Z, Zhang K, Wang Z, et al. 3D pose estimation of large and complicated workpieces
based on binocular stereo vision. Applied Optics, 2017, 56(24): 6822-6836.
Xu P, Ding X, Wang R, et al. Feature-based 3D reconstruction of fabric by binocular
stereo-vision. The Journal of The Textile Institute, 2016, 107(1): 12-22.
Wang Y, Wang X. On-line three-dimensional coordinate measurement of dynamic binocular
stereo vision based on rotating camera in large FOV. Optics express, 2021, 29(4):
4986-5005.
Xue T, Xu L, Zhang S. Bubble behavior characteristics based on virtual binocular stereo
vision. Optoelectronics Letters, 2018, 14(1): 44-47.
Read J C A. Stereo vision and strabismus . Eye, 2015, 29(2): 214-224.
Lemme N J, Li N Y, Kleiner J E, et al. Epidemiology and video analysis of Achilles
tendon ruptures in the National Basketball Association . The American Journal of Sports
Medicine, 2019, 47(10): 2360-2366.
Losada A G, Therón R, Benito A. Bkviz: A basketball visual analysis tool . IEEE computer
graphics and applications, 2016, 36(6): 58-68.
Panagiotakis E, Mok K M, Fong D T P, et al. Biomechanical analysis of ankle ligamentous
sprain injury cases from televised basketball games: understanding when, how and why
ligament failure occurs . Journal of science and medicine in sport, 2017, 20(12):
1057-1061.
Shih H C. A survey of content-aware video analysis for sports . IEEE Transactions
on Circuits and Systems for Video Technology, 2017, 28(5): 1212-1231.
Wang L, Xiong Y, Wang Z, et al. Temporal segment networks for action recognition in
videos . IEEE transactions on pattern analysis and machine intelligence, 2018, 41(11):
2740-2755.
Chen C, Jafari R, Kehtarnavaz N. A real-time human action recognition system using
depth and inertial sensor fusion . IEEE Sensors Journal, 2015, 16(3): 773-781.
Shahroudy A, Ng T T, Gong Y, et al. Deep multimodal feature analysis for action recognition
in rgb+ d videos. IEEE transactions on pattern analysis and machine intelligence,
2017, 40(5): 1045-1058.
Ding S, Qu S, Xi Y, et al. Stimulus-driven and concept-driven analysis for image caption
generation . Neurocomputing, 2020, 398: 520-530.
Yu N, Zhai Y, Yuan Y, et al. A bionic robot navigation algorithm based on cognitive
mechanism of hippocampus . IEEE Transactions on Automation Science and Engineering,
2019, 16(4): 1640-1652.
Amor B B, Su J, Srivastava A. Action recognition using rate-invariant analysis of
skeletal shape trajectories . IEEE transactions on pattern analysis and machine intelligence,
2015, 38(1): 1-13.
Yao Q. Adaptive finite-time sliding mode control design for finite-time fault-tolerant
trajectory tracking of marine vehicles with input saturation . Journal of the Franklin
Institute, 2020, 357(18): 13593-13619.
Demir B E, Bayir R, Duran F. Real-time trajectory tracking of an unmanned aerial vehicle
using a self-tuning fuzzy proportional integral derivative controller . International
Journal of Micro Air Vehicles, 2016, 8(4): 252-268.
Zhang Q, Mills J K, Cleghorn W L, et al. Trajectory tracking and vibration suppression
of a 3-PRR parallel manipulator with flexible links. Multibody System Dynamics, 2015,
33(1): 27-60.
Bouakrif F, Zasadzinski M. High order iterative learning control to solve the trajectory
tracking problem for robot manipulators using Lyapunov theory . Transactions of the
Institute of Measurement and Control, 2018, 40(15): 4105-4114.
Podder, P.; Das, S.R.; Mondal, M.R.H.; Bharati, S.; Maliha, A.; Hasan, M.J.; Piltan,
F. LDDNet: A Deep Learning Framework for the Diagnosis of Infectious Lung Diseases.
Sensors 2023, 23, 480.
Hasan MJ, Sohaib M, Kim J-M. An Explainable AI-Based Fault Diagnosis Model for Bearings.
Sensors. 2021; 21(12):4070.
Hong Liu, a teacher of Henan Light Industry Vocational College, graduated from Henan
University with a major in physical education. During her teaching period, she wrote
and published nearly ten papers and participated in the compilation of four textbooks.
Participate in department-level projects, preside over provincial projects and school-level
projects. In terms of teaching and scientific research, it is the main force of the
college.