Occluded Pedestrian Detection That Combines Multiple Classifiers In Street Scene

With the development of artificial intelligence and deep learning technology,object detection,recognition,tracking and other computer vision applications have been widely used in our lives.Pedestrian detection has gained more and more atten tion in recent years because of its wide application field.Due to the pedestrian’s own rigid and the complexity of the environment,pedestrian detection under various scenes is still the difficulty and focus of the object detection field.Firstly,this paper introduces the research background and significance of pedestrian detection,and illustrates the application of pedestrian detection in our life.Then,this paper introduces the technical difficulties and current research status in pedestrian detection field,and introduces the commonly used features and classification methods in detail.The main work of this paper is to solve the occlusion pedestrian detection b y cascaded multiple classifiers.Training common pedestrian detection BING module instead of sliding window detection method.The main work of this paper can be divided into the following five aspects:1、Based on HOG features and linear SVM classifier,select the appropriate samples from the INRIA dataset.Training a classifier for detecting pedestrian without occlusion.This classifier is in order to test whether different parts of the pedestrian is occluded will lead to the same result.Part of the testing data comes from the CVC-05 part occlusion dataset,and the others are from the internet.2、We made an exclusive occluded pedestrian dataset by extracting positive examples and negative examples from INRIA 、CVC05-PartOcclusion dataset and another dataset that we collected from the Internet.We use HOG and linear SVM to train the rest of eight classifiers,each classifier is used to detect different types occluded pedestrians.There is one classifier for pedestrians without occlusion,two classifiers for leg occlusion pedestrians,three classifiers for each of left and right half body occlusion pedestrians.3、In the detection section,we set the classifiers to two stage.The first stage contains only one classifier,the second sta ge contains eight classifiers.Classifiers between the two stages are cascaded,the second-stage classifiers are parallel.When a candidate window’s score is higher than the threshold value that we set in the first-stage classifier,this window can be brought to the second-stage classifiers for further testing.By using cascaded implementation,our method is as fast as the traditional HOG method.4、A NMS+Merging approach is proposed to fuse multiple classifiers detection results.First,use NMS to remove low score detection window,then different classifier’s retained detection window s using a fusion way and get the final detection results.5、Using BING algorithm replaces sliding window detection method.Based on Caltech dataset training the BING detection template,and the parameter setting is according to pedestrian specific aspect ratio.Our method can shorten the time consumed in the detection stage while keeping the original detection precision.