A vehicle detection and tracking method for traffic video based on faster R-CNN

In this paper we present a vehicle detection and tracking method for traffic video analysis based on deep learning technology. Indeed, with the rapid development of deep neural networks, vision-based approaches for vehicle tracking by detection have significantly advanced compared to existing approaches. Therefore, the proposed method is composed of three deep neural networks: Feature Network, Region Proposal Network (RPN) and detection network. The Feature Network is used to pre-train and convert video frame to feature maps using a specific convolutional neural network. The RPN network is a an additional convolutional neural network that slides on the feature map and provides a set of bounding boxes that has high probability of containing any object. Finally, a detection Network based on Region-based Convolutional Neural Network (R-CNN) is in charge of assigning a class and bounding box to each region of interest. The main idea of object tracking is to use region of interest for frame-by-frame tracking by extracting features from the current frame then using object detection from the previous frame to regress their detections in the current frame. Experiment results prove that the proposed method provide a high accuracy rate compared with existing methods.