Pedestrian Detection Algorithm With Multiple Feature And Cascade Classifier

Human detection is a hot research area in the field of pattern recognition and computer vision, and with a strong theoretical significance and application value in fields such as robot vision, intelligent video surveillance and human action analysis etc. At present, human detection algorithm has been obtained some achievement. But those algorithms can’t balance real-time, robustness and accuracy on account of occlusion, complex background and a variety of human pose and appearance. Concerning this issue, proposed a fast and robust human detection algorithm with multi-feature and cascaded classifier. It’s divided into hypothesis generation stage and hypothesis validation stage. The main work and innovation points are as follows:1. The overview of the conception and basic procedure of human detection. Human detection with feature description is introduced in details. We also analyze the advantage and disadvantage of those algorithms.2. In the hypothesis generation stage, due to multi-scale orientation feature can’t describe the human edge contour feature very well, we improve the multi-scale orientation feature by adding two feature blocks based on the original feature block, called "extend multi-scale orientation". And the number of windows which are used to extract human feature decreased from 8 to 2 in order to reduce computation complexity. Then we train coarse classifier by Adaboost algorithm to choose the area which may exist human.3. In the hypothesis validation stage, design a feature to describe the human feature, which called multi-scale globe histograms of oriented gradient. And due to high dimensions and computation complexity of the feature, we take winner take all hash(WTA) into the feature coding in order to sparse feature vector. This method can remove redundant information, and speed up the computation. Then we train fine classifier by Intersection Kernel Support Vector Machines to locate human accurately.4. By experiment on INRIA and TUD-Brussels public test set and comparing with other state-of-art algorithms in the world in order to verify effectiveness and expandability of our method.