Research Of Pedestrain Classification Methods Based On Hierarchical Deep Learning

In security monitoring, traffic safety and other fields, pedestrian detection is important application. In "Car Era", the problem of reducing the rates of road traffic injuries and accidents is deeply concerned by the society. To timely realize whether there is pedestrian or not in front of or near the vehicle, technical studies of pedestrian detection are doing for security warning, which has great significance to reduce and avoid collision between pedestrian and vehicle. For complex environmental factors and the pedestrian itself particularity, we design and imcomplent two pedestrain classification methods. The results with kindred methods on 4 datasets show the effectiveness of our methods for pedestrian classification tasks.The main content and the main innovation of this paper are:(1) For traditional manual features are difficult to adapt pedestrian classification task under complex scenarios, we propose a hierarchical feature extraction method based on unsupervised learning, use the convolution network model to extract feature. The method firstly uses a forward prediction function to train optimal sparse codes, to respectively learn the two-level models with CPSD algorithm under the framework of deep convolution network model, then fusing the features of two layers by method presented in text. Finally, we carry out classification with SVM algorithm.;(2) For lighting conditions, body posture changes and occlusion problem and so on, we puting forward a two-layer image representation method based on sparse coding. Two different sparse representations are obtained from a two-layer sparse coding model, and then to get better sparse representation with an effective feature fusion, fusing traditional features with unsupervised learning features well, producing image descriptor. Then constructing the spatial pyramid histogram representation of the image;Our research work and results have a certain role and reference value for improving the theoretical level and function of a pedestrian detection system.