Context-Aware Object Region Proposals for Efficient Vehicle Detection from Traffic Surveillance Videos Using Deep Neural Networks

摘要

Recently, many methods based on deep neural networks have been developed for object recognition, which dominate various performance competitions on public datasets such as ImageNet and Pascal VOC. Existing methods suffer from high computational complexity and/or insufficient recognition accuracy for practical use. In this paper, we demonstrate that, in specific application domains, such as traffic video surveillance, the priori knowledge or environmental context information can be utilized to dramatically reduce the computational complexity and improve the object detection performance. Specifically, our method models the traffic scene background, using the model as a context to guide the generation of a much smaller number of high quality object region proposals that maintain 100% coverage. We then train a deep convolutional neural network (DCNN) to classify these proposal regions and have achieved 99% accuracy on a large test dataset, which outperforms existing methods DCNN-based methods, such as YOLO.