Pedestrian Detection Based On Deep Learning

Pedestrian detection has become an important branch in the field of computer vision because of the particularity of it’s wide application. It has broad significant applications such as video surveillance, driver assistance system, intelligence robot and so on. For example, the concerns in video surveillance are focused on the pedestrian tracking and the consequent behavior recognition. The former is an important prerequisite for the latter. However there are much differences between pedestrians in dressing and body postures, not to mention that backgrounds are complex, illumination change a lot and occlusion often appeared in the real word. All this challenges making pedestrian detection exiting many problem need to be solved.How to design a good feature which has huge differences from other categories and has little difference under the same category and how to train a good classifier become an important research topic in the field of pedestrian detection. The histogram of oriented gradient (HOG) proposed in2005is an important breakthrough in this field. Consequently most of the pedestrian features are improved based on it. Generally speaking manual design of a perfect feature is very hard to achieve even for researchers who have solid professional knowledge. Fortunately, deep learning can be a good solution to this problem because it can learn features form the data automatically according to different classification tasks.On the basis of intensive investigation on the related technologies of pedestrian detection and deep learning, this thesis summarizes the main difficulties and existing problems in the field of pedestrian detection. The main research achievements are summarized as follows:(1)To avoid traditional requirements of manual design features and improve the robustness of features, this thesis leverages deep learning for pedestrian detection. It can make full use of the advantage of deep convolutional neural network and extract features form the database of pedestrian detection. However, the layer of the deep learning network architecture is often very deep and requires to learn many parameters. We can effectively avoid over-fitting problem when training the network only on the condition that the training data is sufficient. This thesis propose to expand the database through content-based image retrieval as well as taking into consideration of the factors of pedestrian resolution, background distribution and so on. This strategy of expansion can keep the original distribution of INRIA database which can benefit the training of a deep learning network which is more adapted to the INRIA database.(2)To solve the problem of too much redundant windows with poor quality generated by traditional methods, we propose a method of multi-strategy window selection. In this method, we use the selective search algorithm to extract windows which contains high quality windows first. Then make use of normed gradients to filter a large number of redundant windows. At last, we get a smaller number of windows with high-quality, which is of great importance for the subsequent classification task.(3)Based on the above research, this thesis design a pedestrian detection system based on deep learning method. We carried out multi-sets of comparison experiments in this system. After expanding the training dataset by the method of image retrieval, we found that the pedestrian detection system based on deep learning which is trained by the new database can achieve good results under the same false detection rate10%:missing rate is only38%. It gains reduction of8%compared with46%of HOG features. After further combination of our proposed multi-strategy window selection method, the missing rate can be greatly reduced to23%, which gains obvious reduction of23%compared with that of HOG features.