Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Sportifs. --- Volley-ball.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Schools --- Volley ball --- Adolescence

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

ACHIEVEMENT --- VOLLEY BALL --- TRAINING PROGRAMS

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

LEARNING --- TRANSFER (PSYCHOLOGY) --- VOLLEY BALL --- TEACHING

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Nowadays, football is by far the most popular sport in the world. Its huge audience obviously attracts many fields, including the field of game analysis. In recent years, deep learning has brought progress to automated match analysis but a challenge remains strenuous: the ball tracking on football sequences. Indeed, the ball is often represented by few pixels on the football videos and many other object have a similar appearance. In addition, the ball can be occluded by a player or in front of a line with similar color. All these circumstances make the ball tracking a challenging objective.
In this work, an offline ball tracking system based on deep learning for soccer sequences is implemented.
The system is divided into two main parts: a ball detector followed by a ball tracking method. The detector is based on a deep convolutional neural network for image segmentation. A dataset composed of semi-synthetic images representing a synthetic ball in front of a real frame from a football video sequence is created to train the network. The objective of the detector is to identify ball candidates on each frame of the sequence.
Then the tracking method treats all the candidates of the sequence to generate candidate trajectories with the exploitation of the Kalman filter. A score is assigned to each trajectory thanks to the use of a deep neural network for image classification. The trajectories with the highest score are selected as ball trajectories. Finally, a cubic spline interpolation estimates the ball positions between the trajectories.
Evaluated on sequences of professional competitions, this ball tracking system allows an identification of the ball position on more than 80% of the images of most of the evaluated sequences, without having any false detection. Moreover, the evaluation highlights the ability of the system to estimate the ball positions during occlusion on a few frames.

Deep Learning --- Ball Tracking --- Neural Network --- Ball Detector --- Kalman Filter --- Ingénierie, informatique & technologie > Sciences informatiques