一种基于Faster R-CNN的车辆检测算法

A Vehicle Detection Method Based on Faster R-CNN

针对传统的车辆检测算法无法自适应地完成在复杂场景变化下提取目标相应特征的现象,提出了一种基于深度学习的车辆检测算法,该算法结合了Faster R-CNN开源框架和Loc Net网络算法。首先,利用RPN算法获得图片中的候选区域,以减少检测过程中对每张图片的计算量;然后,进入Fast R-CNN网络,利用该深度网络中的卷积层和池化层,自适应地获得车辆目标的所有特征;最后,进入Loc Net网络,通过输入已经得到的图片候选区域,通过卷积层和池化层,不断计算候选区域边界的概率,达到不断优化候选区域边界,最后得到车辆目标的边界框。使用深度学习卷积神经网络,可以避免人工设计车辆目标特征适用性不广泛的缺点,提升车辆目标检测和定位的准确性。

Aiming at the problem that the traditional vehicle detection algorithm cannot adaptively carry out the corresponding feature under the complex scene change,this paper proposes a vehicle detection algorithm based on depth learning,which combines the Faster R-CNN open source framework and the LocNet network algorithm.First of all,the use of RPN algorithm to obtain the candidate area in the pictures,which can reduce the detection process for each picture of the amount of calculation;then entering the Fast R-CNN network,the convolution and pooling layers in the depth network can be used to obtain the characteristics of all vehicle targets in the picture,which can adaptively obtain all the features without the need for artificial design features;finally,after entering the Loc Net network,through the input of the image candidate areas,through the convolution layer and pool layer,which can continue to calculate the probability of candidate region boundaries,to continuously optimize the candidate area boundaries,and finally get the vehicle target bounding box.In conclusion,the depth of learning convolution neural network can avoid the shortcomings of manual design of vehicle target characteristics.As the traditional method of vehicle target detection cannot get higher accuracy,the method in this paper can improve the accuracy of vehicle target detection and positioning.