Traffic Road Detection Based on Dynamic Anchor Frame

Abstract

In recent years, deep convolution neural networks have made great progress in object detection tasks. Generally speaking, the bounding box and the type of bounding box play a very important role in object detection. However, it is not easy for convolution neural networks to directly generate disordered bounding boxes. A widely used solution is to adopt the idea of divide and conquer and introduce the concept of anchor box. At present, anchor frame mechanism has been widely used in top-level object detection framework, and has achieved good results on common datasets. The innovation of this paper is that a novel anchor frame generation method is proposed, which can generate error frames with various aspect ratios for object detection frames. Different from the previous method of generating the anchor box in a predefined way, the anchor box in this method is dynamically generated by the anchor box generator. The feature is that the anchor box generator is not fixed, but learns from anchor boxes defined by fixed rules, which means that the anchor box generator can be adapted to a variety of scenarios. In this paper, the dynamic anchor frame method is used to detect the traffic road. In addition, the weights of the anchor box generator are predicted by a small network whose inputs are predefined anchor boxes. Compared with the traditional anchor frame generation methods, the proposed anchor frame generator has the following innovations: (1) it adaptive adjusts the size and aspect ratio of the anchor frame to improve the quality of the anchor frame. (2) The adaptive IOU country value is used to balance the number of positive samples of the size target. Finally, good efficiency and results are obtained.