Vehicle Detection And Tracking Based On Optical Flow

Recently, vehicle detection and tracking based on video processing is a main research direction of Intelligent Transportation. Optical flow can be defined as a velocity expression method in the imaging plane and it plays a key and basic role in video-based motion analysis because of its ability to obtain the detail 2D motion information of the moving targets between the frames. But the problems which exist in the process of the estimation such as ill-posed problem,high computational complexity are impeding the promotion of its application. Therefore, it is very necessary to find a way that not only can obtain the optical flow in high precision but also in real-time. It will be of great benefit to promote the application of optical flow and the development of Intelligent Transportation.In this paper, the basic principle of differential optical flow method is elaborated, and accordingly, a multi-frame optical flow model is constructed by introducing a prior frame; The linearization process of the energy function is being simplified by using IRLS; GPU is chosen as the implementation platform of the algorithm to achieve real-time operation; A vehicle detection and tracking scheme is proposed based on the real-time optical flow which estimated from the GPU platform.First, despite the fact that temporal coherence plays an important role in processing video data, this concept has hardly been exploited in optical flow methods. A multi-frame optical flow model can be constructed by adding a temporal coherence term. The new model not only can estimate the multiple adjacent flows, but it can also improve the precision of estimation greatly through harder constraints. Meanwhile, the method of iterative reweighted least squares is used to solve the model which can simplify the derivation of the nonlinear energy equations.Secondly, a parallel computing scheme based on graphic processing unit(GPU) is proposed according to the characteristics of optical flow algorithm. The scheme can estimate the optical flow value of several pixels simultaneously under compute unified device architecture(CUDA) and thus can reduce the running time significantly. Experimental results show that, in the case of equivalent accuracy, the running time of the algorithm which is implemented on the GPU is much less than the running time required by the CPU. For the 640*480 image sequences, the processing speed can meet the general realtime application requirements.Finally, an advanced vehicle detection and tracking scheme is proposed base on the real-time optical flow. On the one hand, compared to feature points matching method and the partial dense optical flow method, the advanced scheme not only can obtain more accurate global motion information of the moving targets but can also achieve better moving target extraction result. On the other hand, the advanced scheme can forecast and match the position of vehicle based on the optical flow vectors, also, it can judge and well deal with the common change of the vehicle status during the tracking. The scheme achieves expected results in the experiment.