关键词: 感兴趣层, 非相邻和相邻特征（NNNF）, 随机池化, 深度稀疏自编码

Abstract: Aiming at the problem of slow convolution speed, weak noise, and large redundancy in pedestrian detection in traditional convolutional neural networks, a method based on deep sparse auto-encoder network (DSAEN) is proposed. Firstly, add an interested layer after the input layer, enrich the symmetry characteristics of the image based on the nor-neighboring and neighboring features (NNNF). At the same time, fuse multi-channel cross channel features, like the LUV color space and the gradient amplitude channel. Utilize the improved stochastic pooling method in the regional processing of nor-neighboring feature designs to eliminate the influence of pedestrian deformation, so that the main features that represent pedestrian information are obtained. Then use four auto-encoders to design the four-layer hidden layer depth network. The cross-entropy loss function, the improved ReLU activation functions, and the sparse representation theory are introduced to construct a new objective function, from which the network is learned and the internal structure of the data is searched for. Finally, the classifier is trained using the effective features of the fourth hidden layers. Experimental verification on the public database shows that compared with other existing methods, this proposed method reduces the average miss rate and running time and has good robustness.

Key words: interested layer, nor-neighboring and neighboring features (NNNF), stochastic pooling, deep sparse auto-encoder