基于卷积神经网络的前向车辆检测方法研究

Forward Vehicle Detection Based on Convolution Neural Network

提出了一种基于卷积神经网络(CNN)的前向车辆检测方法。首先,采用车底阴影和车宽约束条件进行假设生成,针对存在较多具有边缘、对称性等特征物体干扰时产生灾难性计算量的问题,提出了阴影特征结合车宽约束条件判定方法生成车辆假设,排除大量非车底阴影区域,提高了检测效率;然后采用CNN对产生的车辆假设进行验证,建立车辆样本库,训练并对比分析了多种结构CNN,验证本文设计结构能够快速准确地验证车辆假设,对GTI数据库的检测正确率达到98.65%;最后经强弱光照、树荫等多场景实验验证表明,本文方法能够快速准确地检测出前方车辆。

A forward vehicle detection method based on vehicle width constraint and convolutional neural network(CNN)was proposed.Firstly,we used car shadow and vehicle width constraint to generate hypotheses Aiming at problems that there will be catastrophic calculations when existing too much interference with Edge,symmetry,texture and so on,the vehicle shadow and width constraint decision method was proposed which eliminated a large number of non-vehicle shadow regions and raised the efficiency of detection;Then,we verified the generated vehicle hypotheses by CNN.Through comparisons of different structures of CNN,the proposed CNN structure was proved to be the best,whose accuracy was 98.65%towards GTI database;Lastly,results of experiments under different conditions showed that,the proposed method was effective in detecting forward vehicles accurately.