基于卷积神经网络的道路车辆检测方法

摘要

A Convolution Neural Network (CNN) based vehicle detection method was proposed.First,an edge enhancement-based road detection and an adaptive shadow segmentation approaches are put forward to resolve the grayscale variation on the road and reduce the influence of the lighting variation.Then,the CNN structure applied to the road traffic environment is determined to train the established sample sets.On the basis of this,the vehicle region,which is wrongly detected as the vehicle shadow,is recognized by CNN and removed from the preliminary detection results,thus,the final vehicle shadow is obtained.Finally,CNN is modified to three classification to verify the advantages of a strong expandability of this method.The experimental results show that this method is effective in various conditions and meets the accuracy requirement,decreasing the false positive rate.%提出了一种基于卷积神经网络的前方车辆检测方法.首先,根据车底阴影特征,运用基于边缘增强的路面检测算法以及车底阴影自适应分割算法来分割并形成车底候选区域,以解决路面灰度分布不均及光照条件变化问题;其次,运用针对道路交通环境的卷积神经网络结构,建立图像样本库进行网络训练;在此基础上,采用基于卷积神经网络识别的方法以验证并剔除被误检测为车底阴影的候选区域,进而确定真正的车辆目标;最后,修改网络为三分类识别,以验证本文方法的强扩展性的优势.实验结果表明:本文提出的车辆检测方法能够很好地区分车底阴影和非车底阴影干扰,有效地提高车辆检测的准确率和可靠性,降低误检率.