Research On Pedestrian Detection Based On Classifier Algorithm

Pedestrian detection in videos and images is one of the most important subjects in the field of computer vision. The wide range of applications of pedestrian detection intersects with many aspects of our lives: automatic driving, surveillance system and military reconnaissance. The underlying intellectual challenges of pedestrian detection have attracted many researchers in academia and industry. Determining the moving foreground of pedestrian, describing the pedestrian in the images and constructing the pedestrian classifier should be all researched for good pedestrian detection results.At present, our predecessors have had some achievements on pedestrian detection. In some simple scenes, the previous pedestrian detection methods can achieve good effect. But those methods do not perform well in some more complex scenes, especially in the case of dynamic background and partial occlusion.We make contributions to determining moving foreground,designing features and improving classifier to promote the effect of pedestrian detection, and make some progress in the situation of dynamic background and partial occlusion.Firstly, the method we advanced that has excellent performance in moving foreground. It is based on two autoencoders. One is to extract the background and the other is to learn. We improved the construction of the network and added random noise to the network. This newer method is faster and more robust.Secondly, some comparisons between several kinds of pedestrian features we made show that CENTRIST is good and simple enough to describe pedestrian in videos and images. The Union-CENTRIST is constructed based on human parts. Experiments show that Union-CENTRIST has better invariability than the original CENTRIST.Thirdly, we propose a method that combines multi-autoencoders and SVM to construct the pedestrian classifier. The multi-autoencoders is to keep the independence of each parts of the Union-CENTRIST. Every autoencoder is designed for the specific part of human to encode the component of the part. Each component of Union-CENTRIST is compressed and the essential information of the feature is kept. We combine the ideas of different classification algorithms and the experiments show that we achieved better pedestrian detection effect.