基于YOLO的驾驶视频目标检测方法

Object Detection on YOLO for Vehicle Camera Videos

运用YOLO(You Only Look Once)实时目标检测算法解决了驾驶视频目标检测问题。针对目标检测算法受环境条件影响鲁棒性差、小目标识别能力不高的问题,建立了涵盖多种天气环境、包含疑难目标的驾驶视频样本数据库,提出了疑难样本训练方法,训练出可在多种天气环境中良好识别小型汽车、行人、公交车及货车的YOLO检测模型。实验结果表明,该训练方法可有效提升目标检测性能;所得检测模型具有较高的召回率和精确度,可初步应用于实时驾驶视频的目标检测。

Applying YOLO(You Only Look Once)real-time object detection network to detection tasks in vehicle camera videos.Aiming at the low robustness due to environmental impact,and low performance of small object detection of present algorithms,a dataset containing multiple weather condition samples and Difficult Samples is established.The method of training Difficult Label is proposed.A YOLO model capable of well recognizing small cars,pedestrians,buses,and trucks in complex driving environment is trained,with high Recall rate and Precision,which can effectively improve detection performance and be primarily applied to real-time detection in vehicle camera videos.