一种基于YOLOv3的道路多目标检测方法

A Road Multi-Object Detection Method Based on YOLOv3

摘要: 道路场景中多目标检测对车辆的自动驾驶和辅助驾驶的智能化有着重要意义。现有道路目标检测算法存在检测精度低、实时性较差和目标漏检等问题。针对这些问题，本文构建一种基于YOLOv3的高精度、低时延、低漏检的道路多目标检测方法。通过对YOLOv3目标检测原理进行深入分析，基于迁移学习(Transfer Learning)的思想，在经过预训练的模型上仅使用Pascal VOC 2007中道路场景常见类别数据对模型进行训练，通过调整学习策略，利用较小的训练集和较少的训练轮次可以获得实时性强、精度较高的目标检测模型，单张图片检测时间只需0.04秒，在测试集上mAP (mean Average Precision)达到了91.5%，实验证明本文方法的有效性，该方法在精度、时延和漏检方面取得了较好的效果。

Abstract: Multi-object detection in road scenes is of great significance for automatic driving and intelligent driving assistance of vehicles. The existing road object detection algorithms have some problems, such as low detection accuracy, poor real-time performance and missing target detection. To solve these problems, this paper constructs a road multi-object detection method based on YOLOv3 with high precision, low delay and low omission. Through indepth analysis of the YOLOv3 object detection principle, based on the idea of Transfer Learning, the YOLOv3 model was trained on the pretrained model using only the common category data of Pascal VOC 2007 (Pascal Visual Object Classes 2007) in the road scenes. By adjusting the learning strategy and using smaller training set and fewer training epochs, a target detection model with strong real-time performance and high accuracy can be obtained. The single image detection time is only 0.04 seconds, and the mean Average Precision on the test set reaches 91.5%. The experimental results show that the proposed method is effective and achieves good results in precision, delay and omission detection.