基于多通道稀疏卷积神经网络的行人检测算法

Pedestrian detection algorithm based on multi-channel and sparse convolution neural network

针对传统卷积神经网络(CNN)的输入是原始图像,冗余信息多,对局部边缘和纹理的刻画不明显的问题,提出了一种基于多通道输入的稀疏卷积神经网络(MCS—CNN)检测算法。将图像方向梯度直方图(HOG)特征构成的HOG特征图和色差(YUV)颜色空间组成3个通道,通过卷积层提取特征,并采用稀疏自动编码器稀疏化;使用Softmax分类器进行行人检测。模型充分利用图像的像素级特征,同时还融合HOG对于行人轮廓显著描述的优点。实验结果表明:与CNN,HOG结合支持向量机(HOG—SVM)检测算法相比,MCS—CNN模型检测准确度和检测速度均明显提高。

Aiming at problem that the input of the traditional convolutional neural network（ CNN） is original images with redundant information,which can＇t describe the local edge and texture characteristics obviously,a multi-channel sparse CNN（ MCS—CNN） detection algorithm is proposed. Histogram of oriented gradient（ HOG） —feature map and YUV color space are formed into three channels,and the features are extracted from convolutional layer,the features are sparsed by sparse Auto Encoder. The softmax classifier is used to detect pedestrian. In addition to taking full advantage of pixel level feature,the model also fuses the advantages of HOG with pedestrian profile description. Experimental result shows that the accuracy and detecting speed of MCS—CNN model is obvious improved,compared with traditional CNN and HOG—SVM detection algorithms.