Abstract

Traffic accident scene recognition is the first premise of accident analysis and scene reconstruction, so the research on traffic accident scene recognition technology is of great significance. In this paper, an improved Region Proposal Network based on Faster Region-based Convolutional Network (Faster R-CNN) is proposed. Because the network adopts four strategies to improve the Regional Proposal Network (RPN), the network structure in this paper is called Multi-strategy Region Proposal Network (MSRPN). The experimental results show that the mAP value of MSRPN algorithm surpasses the other two target recognition algorithms. At the same time, MSRPN only needs to generate 150 region proposals in each image to obtain the above experimental results. In addition, the algorithm has good performance in small target detection. Especially, the target recognition speed is 6 fps, which is faster than other target detection algorithms. In conclusion, the target recognition algorithm based on MSRPN can be effectively applied to target recognition system.

摘要

交通事故场景识别是事故分析和场景重建的首要前提，因此对交通事故场景识别技术的研究具有重要意义。在本文中，提出了一种基于 Faster Region-based Convolutional Network (Faster R-CNN) 的改进的 Region Proposal Network。由于该网络采用四种策略来改进区域提议网络（RPN），因此本文中的网络结构称为多策略区域提议网络（MSRPN）。实验结果表明，MSRPN算法的mAP值优于其他两种目标识别算法。同时，MSRPN 只需要在每张图像中生成 150 个 region proposal 即可获得上述实验结果。此外，该算法在小目标检测方面具有良好的性能。尤其，目标识别速度为6 fps，比其他目标检测算法更快。综上所述，基于MSRPN的目标识别算法可以有效地应用于目标识别系统。