Abstract:
Object detection with object recognition and bounding box generation plays an important role in the visual perception of autonomous driving. With the rapid development of deep learning, a series of object detection algorithms based on deep convolution networks have been proposed. These vision-based algorithms have been widely used to deal with driving scenes or other complicated situations. In this paper, a novel detection algorithm, single shot multi-box detector(SSD)based on deep convolution network-visual geometry group network(VGG), is proposed to achieve target recognition, semantic annotation, and bounding box selection in driving scene. Moreover, an improved algorithm, termed as single shot multi-box detector with aspect ratio selection strategy(SSD_ARS), is also proposed for object detection in auto-driving. Momentum optimization is used in the gradient descent algorithm and the learning rate reduction strategy is optimized. Furthermore, the aspect ratio selection strategy is improved when generating the priori box. It is shown from the results on different selection strategies of aspect ratio when generating the priori box that the proposed algorithm in this work can effectively improve the accuracy of object detection, especially, the detection accuracy of small objects. Finally, the detection experiments of 9 significant object classes in the real driving scene are made, from which it is seen that the proposed algorithm performs well at near and medium distances. Besides, it is also demonstrated via the experiments on driving video that the detection speed of the proposed algorithm in this work can meet the requirements of real-time detection.