Abstract: In this research, through in-depth analysis of the principle of YOLOv3 algorithm target detection, the algorithm model is trained using the common pedestrian, vehicle, traffic lights and other data in the road scene of the self collected data set. By changing the model input, filtering meaningless objects, and adjusting the learning strategy, a target detection model with strong real-time performance and high accuracy is obtained by using smaller training sets and fewer training rounds. The experiment shows that the average precision of the improved method in the test set reaches 71.58%, 1.94% higher than the traditional YOLOv3 algorithm, and the detection speed of the algorithm reaches 70.04 f/s, which is better than the traditional YOLO series algorithms. At the same time, the method is applied to the vehicle driving dynamic data set, which can achieve real-time detection of road objects in the video.

Key words: Target detection; Deep learning; YOLOv3

摘要： 文章通过对 YOLOv3 算法目标检测原理进行深入分析，使用自采数据集道路场景中常 见的行人、车辆和红绿灯等数据对算法模型进行训练，通过改变模型输入、过滤无意义目标 物，以及调整学习策略，利用较小的训练集和较少的训练轮次获得实时性强、精度较高的目 标检测模型。实验表明，改进方法在测试集上平均精度达到 71.58%，比传统 YOLOv3 算法提 高 1.94%，并且算法检测速度达到 70.04 f/s，较优于传统 YOLO 系列算法。同时将该方法应 用于车辆行驶动态数据集，能够实现针对视频中道路目标的实时检测。

关键词: 目标检测；深度学习；YOLOv3