The Research Of Multiple Object Tracking For Surveillance Video With Deep Learning

With the construction of safe city,the number of surveillance video has grown rapidly.Dealing with video requires a lot of manpower and material resources,so intelligent video analysis system has been paid more and more attentions.With the development of deep learning technology,the machine can have a deeper understanding of the image.The development of hardware technology also improves the processing speed.Intelligent video analysis system will be more widely used.Multiple object tracking is an important aspect of intelligent video analysis system,and has a significant impact on the deeper mining of video information.Machine learning method is used to extract artificial characteristics early and now convolution neural network is used to extract spatial information for object tracking.Considering the temporal domain information has a huge impact on the object tracking,LSTM network is used to study the motion law of the object,convolution neural network and recurrent neural network are used to solve tracking with combining the spatial features and the temporal position information.Image sequences are extracted with a certain interval number of frames from a video to improve the speed with little impact on motion law.Trajectory prediction can solve the problem of object occlusion in tracking.Pipeline is used to realize the parallel execution of CPU and GPU,for making full use of hardware resources and improving the processing speed of the system.The brief process of the whole method is as follows: Image sequences are extracted and Faster R-CNN is used to detect objects from images.The image features,object position and object features are extracted from Faster R-CNN network and made as inputs to the LSTM network to predict the object position at next moment.Then the multi-strategy matching method combining the spatial features and the temporal position information is used to determine which object on tracking is related to the object in the current frame.The experimental results show that the multi-strategy matching method can improve the accuracy of the tracking and has a good ability to solve occlusion.The score on the MOT dataset shows that the method gets a good result on accuracy and speed.