Cross-video Real-time Object Tracking Technology Based On Siamese Network

With the rapid development of technology and the continuous improvement of people’s quality of life,the demand for safety is becoming more and more urgent.With the rapid development of video security surveillance industry,various applications also raise different surveillance technology requirements.Digital and grid-based surveillance gradually turned into intelligent surveillance.Among them,object tracking based on surveillance video is the basic requirement.At present,the conventional object tracking algorithm can better track the target from a single camera perspective.However,single-view tracking has some limitations,such as narrow field of view and less information in practice,and multi-view cooperative cross-video tracking is usually applied.Cross-video tracking scenarios pose many challenges to object tracking algorithms,such as real-time performance,occlusion and changes of scale and appearance.How to improve the object tracking algorithm model to meet the needs of cross-video real-time object tracking technology is an urgent problem to be solved.Based on Siamese neural network,this paper optimizes the cross-video object tracking algorithm model from the aspects of real-time tracking and scale oriented appearance change tracking,and constructs a cross-video object tracking prototype system.The main contributions are as follows:1.A light-weight Siamese network object tracking model MPSiam based on multiplexing convolution is proposed.Aiming at the problems of complex network structure and slow inference speed of the state-of-the-art tracking model,spatial multiplexing convolution and channel multiplexing convolution are constructed,which enhances the information flow between space and channel,reduces network parameters and speeds up the inference speed.Experiments show that the inference speed of MPSiam model on CPU and GPU has been improved,and the speed on CPU has reached the standard of real-time tracking.2.A Siamese network object tracking model Tr MPSiam for scale appearance change based on Transformer is proposed.Aiming at the problem that the state-of-theart network model can not make full use of the temporal information between video frames,a transformer structure is introduced to construct the inference relationship between video frames based on MPSiam network.Aiming at the problem of target scale appearance change in cross-video tracking,the anchor-based RPN network is changed to the anchor-free central point target estimation network,and the scale prediction regression network head is added to improve the sensitivity of the model to scale change.A target template updating method based on dynamic template is constructed,which implemented the robust modeling of target appearance with only a small number of parameters.Experiments show that Tr MPSiam model can make full use of temporal information,improve the tracking accuracy,and can better deal with the changes of scale and appearance.3.A cross-video object tracking prototype system is developed and verified by experiments.Firstly,using thin plate spline function and global unified geographic coordinates,the pixel coordinates are corresponding to the global coordinates.Combined with polygon clipping algorithm and ray method,the location and entry detection of overlapping areas are implemented,and the target handoff algorithm is formed.On this basis,a cross-video object tracking prototype system is developed.Experiments show that the system can track conventional targets in real-time and stably across video scenes.