Research And Implementation Of Moving Object Detection And Tracking Algorithm

Moving object detection aims to find the position of all moving objects in varies of background.In this paper,we realized moving object detection system based on DSP platform.The system has varies of functions contains video capture,video display and moving object detection.It satisfy the command of industry.Object tracking aims to locate the initialized object in the whole video sequence,according to continuity of space and time.Object tracking methods base on deep learning have competitive accuracy compared to traditional tracking method,but these deep learning based methods always have bad performance in speed result to limitation of application to industry.We design a real-time object tracking method base on deep learning.The tracker is designed base on fully-convolutional networks with siamese architecture.We can get the confidence value of all candidates through just one forward passing.During offline training,the network of tracker learn similarity between objects in different frames of video sequence,so the network of tracker does not need to update weights online,these strategies can improve speed of tracker obviously.To handle the problem of drifting in fast motion condition,we use real-time optical flow generator FlowNet to help tracker build motion model of object.Perform peak detection in output response map to find the position of object in this frame.Additionally,we use hard negative mining(HNM)in training stage,HNM improves the weights of misclassified samples in loss computing stage.It makes network optimized better in the condition of imbalance between positive samples and negative samples.Experiments show that our tracker has real-time speed when gets state-of-art accuracy.