Pedestrian detection is an important branch of object detection. Nowadays, it's intensively investigated and becoming a hot topic in the field of computer vision. It could be widely used in smart surveillance, driver assistant systems, motion analysis, advanced human-machine interfaces and so on. Its potential application is very promising. The state of the art is mainly based on machine learning, which extracts features and establish pedestrian model by learning from training samples.We combine Viola's algorithm and Dalal's hog(histogram of gradients) feature and apply them to pedestrian detection. We improve the algorithm from the following aspects: First, we use a cascade classifier instead of support vector machine to greatly improve the detection speed. Second,we adopt both hog feature and haar feature to gain more descriptive ability. Besides, we simplify hog feature as shog feature to save computation time. Third, we improve the weak classifier by real adaboost and look up table. Finally, in order to use high dimensional hog feature efficiently in adaboost, we propose to project the hog feature to one dimension via weighted fisher linear discriminant, then estimate the probability by look up table.Experiments show our method is very efficient. When the false positive rate is 1/10000, our detection rate is about 86% on Inria pedestrian dataset. The running speed is about 2 fps with 640X480 images on a 1.8 GHz CPU while our training time is only about 8 hours. |