基于YOLO算法的车辆违停检测

Illegal parking detection based on YOLO algorithm

车辆图像和视频的违停检测,对于交通管理和社会治理方面具有重要意义。传统的车辆检测利用区域选择算法在候选区域上提取特征,但多样化的环境容易造成算法复杂度较高且检测精度较低。采用端到端的YOLO检测算法,首先,通过提取整张图像的特征,利用多层卷积运算和非极大值抑制法来处理提取的特征,检测目标的准确位置;然后,通过损失函数的反向校正,提高了目标检测的精度;最后,利用检测框和违停区域框之间的重叠度及阈值来判断车辆是否违停。在数据集KITTI和实际道路上的图像中进行实验,验证该算法能够实现对目标的精确检测。

The detection of vehicle violation images and videos is of great significance for traffic management and social governance.In the traditional vehicle detection,region selection algorithms were used to extract features on candidate regions,but diverse environments led to higher algorithm complexity and lower detection accuracy.In this paper,we adopted the end-to-end YOLO detection algorithm.Firstly,by extracting the features of the whole image,we analyzed the features with the multi-level convolution operation and non-maximum suppression method,and detected the exact position of the object.Then,the inverse correction was made in terms of loss functions,which improves the accuracy of object detection.Finally,the degree of overlap and the threshold between the detection frame and the violation zone frame were used to determine whether the vehicle was in violation.Experiments on the dataset KITTI and the images of the practical situations show that the algorithm can achieve accurate detection of the target.