基于Faster RCNN的行人及车辆类型检测

Pedestrian and vehicle type detection based on Faster RCNN

随着汽车数量与日俱增,交通事故的发生频次也在增加,针对车辆类型和行人的检测问题,本文在原始Faster RCNN的基础上,首先使用残差网络RES101代替传统的VGG16网络作为共享卷积层,进行图像特征的提取,然后改变原来的锚框尺寸方案,使用锚框尺寸为4、8、16代替原来锚框尺寸,得到行人及车辆类型检测模型。通过在KITTI测试集上的测试结果表明,使用本文模型平均检测准确率可达86.5%,相比原始Faster RCNN平均准确率提高了3.65%,相比于使用残差网络RES101作为卷积层的Faster RCNN平均准确率提高了2.06%。

With the increasing number of vehicles and frequent traffic accidents,aiming at the detection of pedestrians and vehicle types,based on the original Faster RCNN,firstly,the residual network RES101 is used instead of the traditional VGG16 network as the shared convolutional layer to extract the image features.Then,the original anchor frame size scheme is changed,and the size of the anchor frame is 4,8 and 16 instead of the size of the original anchor frame,and the pedestrian and vehicle type detection model are obtained.The test results on the KITTI test set show that the average detection accuracy of the proposed model is 86.5%,3.65%higher than the original Faster RCNN,and 2.06%higher than the original Faster RCNN using residual network RES101 as the convolutional layer.