Xina Cheng

Particle filter with ball size adaptive tracking window and ball feature likelihood model for ball’s 3D position tracking in volleyball analysis

3D position tracking of the ball plays a crucial role in professional volleyball analysis. In volleyball games, the constraint conditions that limit the performance of the ball tracking include the fast irregular movement of the ball, the small-size of the ball, the complex background as well as the occlusion problem caused by players. This paper proposes a ball size adaptive (BSA) tracking window, a ball feature likelihood model and an anti-occlusion likelihood measurement (AOLM) base on Particle Filter for improving the accuracy. By adaptively changing the tracking windows according to the ball size, it is possible to track the ball with changing size in different video images. On the other hand, the ball feature likelihood enables to track stably even in complex background. Furthermore, AOLM based on a multiple-camera system solves the occlusion problems since it can eliminate the low likelihood caused by occlusion. Experimental results which are based on the HDTV video sequences (2014 Inter High School Games of Men’s Volleyball) captured by four cameras located at the corners of the court show that the success rate of the ball’s 3D position tracking achieves 93.39 %.

Anti-occlusion observation model and automatic recovery for multi-view ball tracking in sports analysis

The 3D position of the ball plays a crucial role in professional sport analysis. In ball sports, tracking of balls precise position accurately is highly required, whose performance is affected by inaccurate 3D coordinates and occlusion problem. In this paper, we propose anti-occlusion observation model and automatic recovery by 3D ball detection based on multiview videos to track the ball in 3D space. The anti-occlusion observation model evaluates each cameras image and eliminates the influence of the cameras in which the ball is occluded. The automatic recovery method detects the balls 3D position by homography relation of the multi-video and generates a new distribution to initiate the tracker when tracking failure is detected. Experimental results based on the HDTV video sequences, which were captured by four cameras located at the corners of the court, show that the success rate of the 3D ball tracking achieves 99.14%.

Real-Time 3D Ball Tracking with CPU-GPU Acceleration Using Particle Filter with Multi-command Queues and Stepped Parallelism Iteration

3D ball tracking is a critical function in manyapplications such as game and players behavior analysis, andreal time implementation has become increasingly importantfor it can be used for live broadcast and TV contents. To reacha high accuracy, algorithms usually are time consuming due toa large set of calculations which is challenging to meet realtime demanding. This paper proposes multiple commandqueues, tactical threads allocation and stepped iterativeaddition to empower such a capacity on the CPU-GPUplatform. Multiple command queues achieves a parallelismbetween tasks in the algorithm. Secondly, the tactical threadsallocation helps mapping the algorithm into GPU andenhances synchronism between threads. And this paperproposes stepped iterative addition to achieve partialparallelism in a sequential operation. This work implements inan Intel Core i7-6700 GPU and AMD Radeon R9 FURY GPU.Tracking speed of our work increases 37.8 times from original431ms to 11.7ms while the success rate of the algorithm retainsover 99%. This result fully meets the requirement of 16.6msper frame for 60fps video real-time tracking.

Motion Vector and Players’ Features Based Particle Filter for Volleyball Players Tracking in 3D Space

Multiple players tracking plays a key role in volleyball analysis. Due to the demand of developing effective tactics for professional events, players’ 3D information like speed and trajectory is needed. Although, 3D information can solve the occlusion relation problem, complete occlusion and similar feature between players may still reduce the accuracy of tracking. Thus, this paper proposes a motion vector and players’ features based particle filter for multiple players tracking in 3D space. For the prediction part, a motion vector prediction model combined with Gaussian window model is proposed to predict player’s position after occlusion. For the likelihood estimation part, a 3D distance likelihood model is proposed to avoid error tracking between two players. Also, a number detection likelihood model is used to distinguish players. With the proposed multiple players tracking algorithm, not only occlusion relation problem can be solved, but also physical features of players in the real world can be obtained. Experiment which executed on an official volleyball match video (Final Game of 2014 Japan Inter High School Games of Men’s Volleyball in Tokyo Metropolitan Gymnasium) shows that our tracking algorithm can achieve 91.9 % and 92.6 % success rate in the first and third set.

Vectorized Data Combination and Binary Search Oriented Reweight for CPU-GPU Based Real-Time 3D Ball Tracking

3D ball tracking is of great significance to sports analysis, which can be utilized to applications such as TV contents and tactic analysis. Some applications require real-time implementation, but a highly accurate tracking algorithm is usually time-consuming. This paper proposes a CPU-GPU platform based particle filter for multi-view ball tracking, including 2 proposals: vectorized mask data combination and binary search oriented reweight. The vectorized masks data combination unites HSV mask and inter-frame subtraction mask into one to reduce memory access time. The binary search oriented reweight helps getting and saving reweighted data with low complexity which could directly be used for binary search. The proposed methods are evaluated by both tracking accuracy and execution time. Experiment is based on GPU, the AMD R9 Fury, and compared to the serial implementation on CPU. The tracking accuracy keeps the same, while the execution time is reduced by a factor of 13.

Mixture particle filter with block jump biomechanics constraint for volleyball players lower body parts tracking

Volleyball player body parts tracking is very important for block or jump height calculation which can be applied to TV contents and tactical analysis. This paper proposes a mixture particle filter with block jump biomechanics constraint based on 3D articulated human model. Using mixture particle filters tracking different body parts can effectively reduce the freedom degree of the human model and make each particle filter track the specific target more accurately. Block jump biomechanics constraint executes adaptive prediction model and likelihood model which can make the particle filter specific for block tracking. The experiments are based on videos of the Final Game of 2014 Japan Inter High School Games of Mens Volleyball in Tokyo. The tracking success rate reached 93.9% for left foot and 93.8% for right foot.

Motion Prejudgment Dependent Mixture System Noise in System Model for Tennis Ball 3D Position Tracking by Particle Filter

In tennis game analysis, the 3D position of ball plays a crucial role in score judgment and player evaluation. When tracking the tennis ball in 3D space, high speed and abrupt motion change of the tennis ball are the main problems which make it difficult to predict the near future course of the ball. Aiming at solving above two problems, we propose a system model based on an elaborated mixture system noise. The mixture system noise consists of general change noise and adaptive abrupt change noise which is dependent on motion prejudgment result of tennis ball. The motion prejudgment method is carried out by the current state of ball and players. The motion of ball is classified into general motion and three abrupt motions, including smash, bounce and hit the net. Experiments based on 13 HDTV video sequences, which were recorded by four cameras located at four corners of the tennis court outside in a cloudy day including two players were used to explore the performance of the proposed method. The tracking success rate is 81.14%, gaining 27.64% improvement compared with conventional work.

3D Global Trajectory and Multi-view Local Motion Combined Player Action Recognition in Volleyball Analysis

Volleyball video analysis is important for developing applications such as player evaluation system or tactic analysis system. Among its different topics, player action recognition serves as an elementary building brick for understanding player’s behavior. Most conventional player action recognition methods have limits in real volleyball game due to the occlusion and intra-class variation problems. This paper proposes a 3D global trajectory and multi-view local motion combined volleyball player action recognition method. 3D global trajectory extracts global motion feature through 3D trajectories, which hides the unstable and incomplete 2D motion feature caused by the above problems. Multi-view local motion gets detailed local motion feature of arms and legs in multiple viewpoints and removes clutter features caused by occlusion problem. Through the combination, global 3D feature and local motion feature mutually promote each other and the actions are recognized well. Experiments are conducted on game videos from the Semifinal and Final Game of 2014 Japan Inter High School Games of Men’s Volleyball in Tokyo Metropolitan Gymnasium. The experiments show the combing result accuracy achieves 98.39%, 95.50%, 96.86%, 96.98% for spike, block, receive, toss respectively and improve 11.33% averagely than the sing-view local motion based result.

Motion State Detection Based Prediction Model for Body Parts Tracking of Volleyball Players

Among sports analysis, tracking of athletes’ body parts becomes a popular theme. Marking positions of body parts on the videos which contributes to TV contents and concrete motion capture of athletes which helps promotion of sports technology make sports analysis a commercially-viable research theme. This paper proposes motion state detection based prediction model to predict the near future motions of players’ arms, band-width sobel likelihood model to observe the shape of human body parts and cluster scoring based estimation to avoid huge error. The motion state detection based prediction model can realize the tracking of players’ high-speed and random motions without templates. The band-width sobel likelihood model can fully express unique shape features of target player’s body parts. And the cluster scoring based estimation utilizes k-means cluster method to divide particle into 3 clusters and evaluate each cluster by scoring in order to prevent huge error from similar noises. The experiments are based on videos of the Final Game of 2014 Japan Inter High School Games of Men’s Volleyball in Tokyo. The tracking success rate reached over 97% for lower body and over 80% for upper body, achieving average 64% improvement of hands compared to conventional work [1].

Event state based particle filter for ball event detection in volleyball game analysis

The ball state tracking and detection technology plays a significant role in volleyball game analysis for volleyball team supporting and tactics development. This paper proposes a ball event detection method to achieve high detection rate by solving challenges including: the great variety of event length, the large intra-class difference of one event and the influence caused by ball trajectories. Proposed state vector covers both the event type and the event period length so that the system model can transits various lengths of event period and predicts event types by volleyball game rules. The curve segmental observation model avoids the tracking error influence to evaluate the event period likelihood by referring neighbouring trajectories of the ball. And according to the standard of the ball event, the feature of the distance between the ball and specific court line are extracted to evaluate the ball event type in observation. At last a two-layer estimation method estimates the posterior state which is a joint probability distribution. Experiments of the proposed method implemented on 3D trajectories tracked from multi-view volleyball game videos shows the detection rate reaches 90.43%.