改进Faster R-CNN算法及其在车辆检测中的应用

Improved Faster R-CNN Algorithm and Its Application on Vehicle Detection

为了更准确地得到符合车辆形态学特征的初始候选框,提出一种基于改进的Faster R-CNN模型的车辆检测算法.首先提取目标框的人工标注坐标值,得到标注框的宽度和高度,然后利用K-means算法对所有框的宽高值进行聚类,得出聚类中心点坐标值后,重新设置RPN的锚盒尺寸及比例,对Faster R-CNN算法的3种尺寸和3种比例加以改进.最后选择轿车、SUV、客车和货车4种车型车辆数据,对改进前后的Faster R-CNN模型进行训练,比较改进前后的模型在车辆检测及车型识别任务中的表现.实验结果表明,使用改进的Faster R-CNN模型达到86.54%的检测准确率,较原始模型提高3.12%.并且该模型有效解决了漏检和误检问题,在恶劣天气和复杂背景下均表现出较高的鲁棒性.

In order to obtain an initial candidate frame that conforms to the morphological characteristics of vehicles more accurately,a vehicle detection algorithm based on the improved Faster R-CNN model is proposed.First,the coordinate values of target frames are extracted to get width and height values of labeled boxes,and then K-means algorithm is used to cluster the width and height values of all boxes.Second,by resetting the anchor box size and the anchor box ratio of region proposal network(RPN)according to the coordinates of the cluster center point,the three sizes and three ratios of the Faster R-CNN can be improved.Finally,vehicle data of four types including cars,SUVs,buses and trucks are selected to train both the unimproved and the improved Faster R-CNN models.At the same time,the performance of the two models in vehicle detection and vehicle identification tasks are compared.Experimental results show that the improved Faster R-CNN model can achieve 84.69%detection accuracy,which is 3.12%higher than the original model.The algorithm effectively improves the missed detection and false detection problems,and shows high robustness in bad weather and complex background.