【摘要】 针对处在恶劣的天气、严重遮挡、过暗或过亮的光照等复杂环境下，现存的目标检测算法对车辆的检测准确度不高，针对该问题提出了基于YOLOv4的改进算法来检测目标车辆。使用图像处理算法处理数据集，模拟复杂环境，以增强算法的鲁棒性；使用K-means++聚类算法优化先验框参数，提高先验框与目标的匹配度；在骨干网中加入空洞卷积（Dilated Convolution）模块，使骨干网络能更好地提取车辆特征；使用Focal Loss损失函数代替交叉熵损失函数，解决检测过程中正样本数和负样本数相差过大的问题。通过实验可得m AP为86.13%，相较原YOLOv4算法提高了7.31%，检测精度在一定程度上优于原YOLOv4检测算法。更多还原

【Abstract】 Aiming at the complex environment such as bad weather,severe occlusion,too dark or too bright light,the existing target detection algorithms have low detection accuracy for vehicles. To solve this problem,an improved YOLOv4 algorithm based on deep learning is proposed to detect the target vehicle. Use image processing algorithms to process data sets and simulate complex environments to enhance the robustness of the algorithm;Use K-means++ clustering algorithm to optimize the parameters of a priori frame to improve the matching degree between the a priori frame and the target;Add Dilated Convolution to the backbone network module,so that the backbone network can better extract vehicle features;Use the Focal Loss loss function to replace the cross entropy loss function to solve the problem that the number of positive samples and the number of negative samples is too large in the detection process. The mAP obtained through the experiment is 86.13%,which is 7.31% higher than the original YOLOv4 algorithm,and the detection accuracy is better than the original YOLOv4 detection algorithm to a certain extent.更多还原