Pedestrian Detection Based On Infrared Images

Pedestrian detection based on infrared image is a technology which is to find and mark pedestrians’ locations accurately from infrared images.As an important and challenging computer vision research field, it is widely used in intelligent transportation, intelligent monitoring, on-board auxiliary, etc. There are two difficultiesin solving the problem. Differentpedestrianheights and different views make different scales of pedestrians, how to quickly find out possible pedestrian candidates is one of the most difficult problems of this technique. The particularity of the infrared images bringabout another difficulty. The pedestrians in infrared images are lack of characteristics such as color, texture. So how to distinguish the pedestrians and non-pedestrianaccurately from the possible pedestrian candidates is another challenging problem.Based on the current problemson pedestrian detection,and after reviewing and analyzing of related literatures which have been presented, we propose two algorithms to solving the pedestrian detection based on infrared image problem. One is based on BING and PCANet, another is based on BING and DPM.In order to figure out how to quickly find out possible pedestrian candidate regions, we use target detection algorithm of BING to find out therough pedestrian proposals. This algorithm use the low level features, the normed gradients as features to train two-layer cascade SVM classifier. The first layer is used to classify the sliding windows, and the second is used to classify the results of the first layer. At last,objectness scores are calculated by the result of second layer.Experimental results indicate that this algorithm can quickly find out possible pedestrian candidate.In order to obtain more precise classification results, we use the global information obtained by PCANet and local information obtained by DPM for judgement, respectively.PCANet use base vectors which are obtained by PCA to replace the filters which are trained in CNN. This algorithm reducesthe training time and the number of parameters of the model, while obtaining the same results as traditional neural network. DPM use part models for representation. It combines the scores of parts and the score of the whole to judge whether a candidate is a pedestrian.Experimental results on the infrared imageindicate that these two algorithm is effective.