A fast and effective video vehicle detection method leveraging feature fusion and proposal temporal link

Vehicle detection in videos is a valuable but challenging technology in traffic monitoring. Due to the advantage of real-time detection, Single Shot MultiBox Detector (SSD) is often used to detect vehicles in images. However, the accuracy degradation caused by SSD is one of the significant problems in video vehicle detection. To address this problem in real time, this paper enhances the detection performance by improving the SSD and employing the relationship of inter-frame detections. We propose a feature-fused SSD detector and a Tracking-guided Detections Optimizing (TDO) strategy for fast and effective video vehicle detection. We introduce a lightweight feature fusion sub-network to the standard SSD network, which aggregate the deeper layer features into the shallower layer features to enhance the semantic information of the shallower layer features. At the post-processing stage of the feature-fused SSD, the non-maximum suppression (NMS) is replaced by the TDO strategy, which link vehicles of inter-frames by fast tracking algorithm. Thus the missed detections can be compensated by the propagated results, and the confidence of the final results can be optimized in the temporal. Our approach significantly improves the temporal consistency of the detection results with lower complexity computations. We evaluate the proposed method on two datasets. The experiments on our labeled highway dataset show that the mean average precision (mAP) of our method is 8.2% higher than that of the base detector. The runtime of our feature-fused SSD is 27.1 frames per second (fps), which is suitable for real-time detection. The experiments on the ImageNet VID dataset prove that the proposed method is comparable with the state-of-the-art detectors as well.