Cooperative Target Tracking Of Multiple Cameras Based On Embedded Platform

In recent years,with the rapid development of monitoring technology,the requirements of intelligent monitoring equipment and the function of monitoring system are increasing gradually.So the video analysis of multiple cameras is becoming more and more popular.However most of the researches still remain in the algorithm and far behind the actual demand.In order to solve this problem,we proposed a multi-target tracking solution with multiple cameras based on ARM embedded platform.In this way,we got a good performance in detection and tracking.Existing traditional intelligent monitoring system mostly based on PC processing platform which has lots of disadvantages such as high power consumption,large volume and poor scalability.In this thesis we first analyzed,chose and constructed the hardware of embedded image processing platform.Linux is chosen as operating system of image processing platform.Then we introduced usage of V4L2,hardware unit of IPU in IMX6Q high-speed image acquisition and transcoding mechanism,which made a solid foundation for the tracking.We optimized Gaussian background modeling method in embedded environment to extract dynamic target for discussion and selection of multi-object.Furthermore we also use HOG features and SVM classifier to detect and recognize pedestrians.After that we can remark pedestrians which will remain in tracking list from all the foreground.In terms of target tracking,we compared Mean-Shift algorithm and Cam-Shift algorithm.Considering real time and implementation effect,we chose Cam-Shift as the basic tracking method.For rapid and stable tracking,we combined Kalman filter theory with Cam-Shift algorithm to improve the tracking efficiency and robustness of the system.Template matching sequence is introduced,the multiple target tracking of single camera has achieved,target's generation and disappearance are also discussed and formulated.We achieved target maching between different cameras by matching one-dimensional pedestrians'color gradient feature with cosine similarity calculation according to the color characteristics of target pedestrian and the luminance difference of cameras.Experimental results show that system tracking success rate reached 80%for dynamic target tracking of suitable size.What's more,the tracking accuracy of multi-objective matching under different angles reached more than 75%.Finally,we summarized the research work of this thesis,and looked into the distance of further research.