The Research On Real-time Tracking In Complex Scenes And Implement On DSP

Tracking is a fundamental problem in computer version,with applications in video surveillance,human-machine interfaces and robot perception etc.Recent years,there are many new methods proposed,but most of those methods have excellent performance only in special target or background.However,those algorithms exhibit limited generalization ability in real world applications.In order to tackle the challenges of tracking deformable objects under moving background and the difficulties of DSST algorithm,a novel tracking method is proposed.The main researching work is as follows:Considering that most tracking methods demonstrate poor performance in tracking moving target with moving background,we first propose a novel background registration based tracking method,which changes moving background to static background among tracking process.This method adopts registration algorithm for the registration of background of scene,which compensates the background shifts and scale changes generated by camera rotation,translation or zoom.After that,advanced static background tracking methods can be employed for achieving high performance object tracking.The proposed algorithm is able to track object in slow changing background,but it cannot satisfy the demands of real-time tracking tasks due to the high expenses of computing registration.We then propose an improved DSST algorithm which greatly reduces the tracking failures of traditional DSST algorithm caused by the accumulation of drift error.This algorithm establishes target model for the related moving target to rectify the drift error among the tracking process of DSST.We then incorporate Kalman filter into DSST,which realizes the optimization of the short-time occlusion problem.Compared with the traditional DSST tracking algorithm,the improved DSST algorithm achieves better performance in both tracking accuracy and success rate.Then,a real-time state machine tracking algorithm is developed based a variety of real-time algorithms which aims at solving the generalization and adaptation ability of tracking algorithms in different scenarios.This algorithm adaptively switches tracking mode based on the analysis of target scale,background complexity,moving target color complexity,which decreases the tracking error of dynamic background.Finally,the high-performance multi-core DSP tracking hardware system is designed for portable and embedded object tracking tasks.The system employs C6678 multi-core DSP processor as the main processor,which executes the tasks of complex arithmetic operation and high performance data processing;FPGA works as a slave processor,to achieve the functions of logic control and character overlay,etc.The designed DSP target tracking system can execute multiple tracking algorithms in parallel fashion,and the developed real-time state machine makes the final decision of the tracking results.