The Real-Time Vehicle Tracking Research Based On The Deep Learning

Intelligent traffic as part of artificial intelligence,in recent years become the focus of research.Among them,the processing of video surveillance has become an important channel to intelligent traffic study.Through the processing of video stream,the location,behavior and safety factor of automobile target are the important work to study intelligent traffic,and the detection of vehicle target is extended to continuous Frame is formed on the movement of the car tracking,the detection of illegal,to combat crime and enhance the contribution of road traffic safety.At the same time,how to locate the target quickly and accurately,and how to track the target becomes the focus of the current research.However,the existing tracking algorithm is simple,the traditional algorithm has several shortcomings: A single algorithm is not working well;partial algorithm improvements are only for one situation,such as: occlusion,deformation and so on.At the same time,accuracy and time complexity can not be balanced.In view of the current research situation,this paper begins with the traditional vehicle target tracking algorithm to carry out understanding of the study,for the traditional Gaussian model easy to reveal the background detection area for the foreground problem and the complexity of the noise treatment effect of complex scenes,the target detection stage This paper proposes a moving object detection algorithm based on improved hybrid Gaussian model.The three frame difference algorithm is used to quickly determine the advantages of background and foreground background,and improve the adaptability of the algorithm to the background.Method to improve the anti-interference of complex background noise.Experiments show that the algorithm reduces the noise in the complex background,short learning cycle,improves the resistance to the weather and the vibration of the camera,and improves the "shadow" noise caused by the background in comparison with the traditional method.Finally,this paper studies the application of depth learning in the field of vehicle tracking: using Selective Search for image segmentation;using a cascade structure to design a two-level cascade structure;low stage classification network has a simple three-layer convolution network and SVM classification;high stage classification network has four-layer convolution neural network structure combined with the SoftMax classifier;using ReLUs to optimize the calculation,to speed up the convergence,easing the over-fitting.Experiments show that the network has achieved good accuracy and reduced the center error,but there are still some problems,such as large time consumption and large demand for training samples.