Model-based Vehicle Tracking using Bayesian Inference Model under Uncalibrated Camera Views

Vehicle tracking is an essential Computer Vision (CV) problem, which can be widely used in many different areas, such as transportation control, security guard, accident detection on highways and so on. Model-based approach has been widely used in object (car) tracking. Some approaches [1]-[3] model the target by a wire-frame model with multiple points. Then the algorithms try to minimize the squared sum of distances from these points to the matching target. For each model feature, a candidate scene feature is located by searching in certain direction, usually a direction that is perpendicular to the projection of the model feature. However, these methods have limitations that they work only when the model features are close to the true matching scene features. To fix this, voting-based schemes for estimating the pose parameters have been proposed [4][5][6]. Though the voting based algorithm is more accurate and robust, it’s time-consuming thus not suitable for the real-time applications. In this paper, we propose a model based vehicle tracking framework under uncalibrated camera views. Two key components, vehicle tracking and vehicle model registration, are included in our proposed framework. In the vehicle tracking part, we implement the FG/BG detecting, Blob detecting/tracking, and motion direction and speed detecting. Then we apply the vehicle model registration scheme based on the detected Blob. Here, a Bayesian inference model is used to fit the vehicle model with the moving vehicle in the each video frame. Compared with existing methods, our method is more robust, since it does not require a calibrated camera. Bayesian inference model can offer the information for correcting the projection of the vehicle model. Moreover, we propose a hidden line removal algorithm, which is more suitable for operating on the vehicle model and reduced the complexity of the traditional hidden line removal algorithms. This paper is organized as follows, In Section II, we provide the design of our proposed vehicle tracking schemes. In Section III, the vehicle model registration based on Bayesian inference model is presented. Finally, we present experiments and results of our proposed scheme and draw the concluding remarks in Section IV and Section V respectively.