Abstract

Road traffic accident scenes provide useful information for understanding how accidents happen and calculating the speeds of the vehicles involved. Unmanned aerial vehicles can obtain photographs of accident scenes, but utilizing these photographs has problems such as low target resolution and scale changes. An improved Resnet–Single-Shot Multibox Detector (R-SSD) algorithm based on a deep residual network (Resnet) is presented to address these problems. A residual network with better characterisation ability is proposed to replace the basic network, and residual learning is employed to reduce difficulty in network training and improve target detection accuracy. The proposed aerial target detection algorithm, based on feature information fusion (I-SSD), addresses the problems of repeated detection and small-sample missed detection in the original SSD target detection algorithm. Based on the detection results, a road traffic accident scene mapping system using either AutoCAD or hand-drawing can be designed.

摘要

道路交通事故现场为了解事故如何发生和计算相关车辆的速度提供了有用的信息。无人机可以获得事故现场的照片，但利用这些照片存在目标分辨率低、尺度变化等问题。提出了一种基于深度残差网络 (Resnet) 的改进的 Resnet-单次多盒检测器 (R-SSD) 算法来解决这些问题。提出了一种具有更好表征能力的残差网络来代替基础网络，并采用残差学习来降低网络训练的难度，提高目标检测的准确性。提出的空中目标检测算法，基于特征信息融合（I-SSD），解决了原有SSD目标检测算法中重复检测和小样本漏检的问题。根据检测结果，可以设计使用AutoCAD或手绘的道路交通事故现场制图系统。