Accurate On-Road Vehicle Detection with Deep Fully Convolutional Networks

Vision-based on-road vehicle detection is one of the key problems for autonomous vehicles. Conventional vision-based on-road vehicle detection methods mainly rely on hand-crafted features, such as SIFT and HOG. These hand-crafted features normally require expensive human labor and expert knowledge. Also, they suffer from poor generalization and slow running speed. Therefore, they are difficult to be applied in realistic application which demands accurate and fast detection in all kinds of unpredictable complex environmental conditions. This paper presents a framework utilizing fully convolutional networks (FCN) to produce bounding boxes with high confidence to contain a vehicle, and bounding box location refinement with SVM to further improve localization accuracy. Experiments on the PASCAL VOC 2007 and LISA-Q benchmarks show that using high-level semantic vehicle confidence obtained by FCN, higher precision and recall are achieved. Additionally, FCN enables whole image inference, which makes the proposed method much faster than the object proposal or hand-crafted feature based detectors.