改进YOLOv4的车辆图像检测算法研究

Research on vehicles image detection algorithm for improved YOLOv4

针对自动驾驶车辆真实行驶场景下因环境复杂,车辆间目标遮挡、环境背景遮挡等导致的车辆检测误检、漏检和定位不准的问题,本文提出了一个改进YOLOv4模型的车辆检测算法。该算法在YOLOv4网络的Backbone与Neck的通道处以及Neck的上采样与下采样处分别添加7处CBAM注意力机制,以提升网络提取有效特征的能力。并利用k-means聚类算法生成适合数据集的锚框。为检验模型的有效性,对数据集进行重新整理与划分,将与车辆无关的种类删去,将Car、Bus、Truck三类合并为Vehicle一类,随后进行实验,并与当前主流的其他目标检测模型进行对比。实验证明,改进的YOLOv4算法比原算法AP提升了4.8%,准确率提升了4.54%,召回率提高了0.9%,优于大部分主流算法。提出的模型为复杂环境下自动驾驶领域的车辆识别提供了有效方法。

For the problem of vehicles misdetection,missed detection and inaccurate localization caused by complex environment,inter-vehicle target occlusion and environmental background occlusion in real driving scenarios of autonomous vehicles,a vehicle detection algorithm with improved YOLOv4 model is proposed in this paper.The algorithm adds seven CBAM attention mechanisms at the channels of Backbone,Neck of the YOLOv4 network and at the upsampling and downsampling of Neck,respectively,to improve the ability of the network to extract effective features.And the k-means clustering algorithm is used to generate the anchor frames suitable for the dataset.In order to test the effectiveness of the model,the dataset is rearranged and divided,the categories unrelated to vehicles are deleted,and the three categories of Car,Bus and Truck are combined into one category of Vehicle,then experiments are conducted and compared with other current mainstream target detection models.The experiments demonstrate that the improved YOLOv4 algorithm improves AP by 4.8%,accuracy by 4.54%,and recall by 0.9%over the original algorithm,which is better than most mainstream algorithms.The proposed model provides an effective method for vehicle recognition in the field of autonomous driving in complex environments.