Application And Research Of Video Object Tracking Algorithm Based On TLD

With the rapid development of the Internet of Things technology, surveillance cameras gradually cover every corner of our daily life and video tracking technology is also used in various practical surveillance scenarios.Object tracking is a hot field of computer vision, and the effect of various object tracking algorithms is different. TLD(tracking-learning-detection) is a robust single target tracking framework, but the result of object tracking is affected by various factors in practice, such as lighting changes, object strain, object rotation and object occlusion and so on. TLD framework based on LK optical flow is more sensitive to track object with deformation and rotation. In order to enhance robustness of the tracking framework in complex environment, Mean shift and particle filter tracking algorithm based on color features are used to replace the LK optical flow and experimental results show that the TLD framework based on mean shift algorithm has better realtime performance and accuracy.TLD framework has good tracking performance in single camera. But in order to achieve a wide range of long tracking, multi-camera system is generally required for object tracking. Due to the need for object handoff and data integration among multiple cameras, the TLD framework can not be directly applied to multi-camera tracking. A kind of non-overlapping multi-camera tracking system is proposed by improving modules of TLD framework. TLD framework maintains a unified sample classifier which uses affine transformation to generate new samples and update classifier parameters through online learning to achieve data integration among multiple cameras. To achieve object handoff, detection module scans the video frame of a certain range of the camera. Then target result is obtained by comparing the similarity. Mean Shift and particle filter tracking algorithm based on color feature are used to replace of TLD framwork's tracking module and experiments are carried out. Experiments result show that the system can achieve continuous tracking in non-overlapping multi-camera and the TLD framework based on Mean Shift tracking algorithm has better robust and accuracy.Through the improving of TLD framework, the paper solves the original framework shortcomings which is not satisfied in the scene of rotation and deformation. TLD framework is applied to multi-camera surveillance system, and it achievs good results.