Pedestrian and Object Detection Using Learned Convolutional Filters

Object detection is still a very active field in Computer Vision. Until now, part based models proved to be one of the most interesting and successful approaches in object and pedestrian detection. The method applies a machine learning approach not to the input images themselves, but to histograms of gradients. However, its performances are still limited when compared to what humans can do. The purpose of the present paper is to show that sparse representations can be successfully used in object detection. The main advantage of using this method is related to the possibility of learning only those filters that are able to express the most frequent patterns that appear in the analyzed images. The experiments are carried out on two widely used datasets, namely VOC2007 and INRIA Person.