Research On Pedestrian Detection Method Based On Sparse Representation

Pedestrian detection is a focus areas of object detection, and increasing number of researchers have extensively studied. There are many important applications about pedestrian detection, such as video surveillance and advanced driver assistance systems. However, it faced with many complex problems, such as pedestrian own posture, varied dress, and also it is highly vulnerable to illumination, complicated background and other external environmental. During to these difficulties, the research of pedestrian detection is full of challenging. The main research in this article is about feature extraction of pedestrian. We provide an effective feature extraction by investigating of sparse representation, thereby improving the accuracy of pedestrian detection. The specific contents in this article as follows:(1) The method of multiple sparse dictionaries histogram for extract pedestrian feature is realized. Pre-learning several different sparse of pedestrian dictionaries through the sparse representation. The pedestrian image is sparse representation by a number of different sparse dictionaries, with obtained corresponding coefficients of the atoms in the dictionary. Statistics each atom in the dictionary for the cumulative histogram, and finally obtained multiple different sparse dictionaries histogram as pedestrian descriptor. The multiple sparse dictionaries histogram has record the statistics information of the pedestrian features. Experimental data indicate that it can be learning enrich pedestrian features by sparse representation with multiple dictionaries. This feature can effectively describe pedestrian and can decrease the feature dimension, while maintaining good detection performance.(2)The method of multiple sparse dictionaries histogram fusions of the histogram of oriented gradients is given. The histogram of oriented gradients recorded the statistical information of the image gradients, while the multiple sparse dictionaries histogram recorded a wealth of statistical information of the image features, such as corners, edges and etc. It can be extract the multiple sparse dictionaries histogram through the sparse representation, then integration of the histogram of oriented gradients as pedestrian descriptor. It can be effectively overcome the problem of insufficient capacity of a single feature, moreover, enhanced the complementary of each other with combining the two features. By experimental data and detection in a real scenario indicate that the fusion features can be more accurate description of the pedestrian and improve the robustness while maintaining good detection performance.(3)The method of feature extraction via sparse representation in the framework of the deformable part models is realized. Divide the pedestrian image into blocks, and extract the sparse dictionary histogram features via sparse representation. Series the histogram features with all the blocks as pedestrian image features, then using the deformable part models for pedestrian detection. Experimental results show that the method can effectively improves the accuracy of pedestrian detection, the natural scene detection also achieved well results.