Research And Implement Of The Key Techniques Of Multi-object Tracking Based On DSP Image Processing System

With a stronger and stronger demand for the intelligent video analysis,the intelligent video analysis is so important that its research is at a hot state.In the intelligent video analysis,the detection and tracking of objects in a video is a pretty important part,which attracts many scholars from inside to outside.Compared with single object tracking,multi-object tracking(MOT)has more realistic value and is more challenging.For producing trajectories with high accuracy,a research about the key technology of MOT is done and a MOT solution is proposed.In consideration of practical significance,the research in this thesis is based on DSP image processing system,and the MOT solution needs to meet the requirement of real time.Then,an image processing system is built with DSP as the system core.In the last,the algorithm is implemented on the image processing system and the code goes through the code optimization.The main contents of the thesis include:1.The popular algorithms about object detection are introduced.After comparison with other algorithms,deformable part model is adopted to perform object detection for providing the following tracking with accurate object information.2.The popular algorithms about MOT are introduced.The “tracking by detection” scheme,that is,data association,is chosen to track multiply objects.Then,network flow is used to model the data association process.An improvement on network flow is made by combining multiply features of an object,and thus the network flow algorithm has been formed.3.To achieve MOT in real time,a research about sliding time window is done and a real-time tracking solution based on the network flow algorithm is proposed.However,it is likely to come out that the overlapped parts of two trajectories from two adjacent time windows are different.To deal with this problem,an effective and reasonable solution for trajectories fusion is proposed.Above all,the MOT solution has been formed.4.The DSP+FPGA framework is adopted and an image processing system is built with TMS320C6678 as the system core.On the system platform,the MOT solution is implemented and the C code achieves optimization.After tested,the system can perform MOT for 640*480/10 fps videos in real time.