Pedestrian Detection Based On Human Body Parts In Complex Scene

Pedestrian detection has a wide range of important application in military &national defense, distress person searching, traffic condition recognizing, the auxiliary driving and intelligent household and so on. This detection based on image characteristic is to locate the pedestrian position in the images and marks out the pedestrians with blank box. The detection algorithm firstly learns learns the characteristic of image to form the detector, then put image into the detector to detect the pedestrians.Because of the uncertainty of the environment scene in reality and the body’s own non-rigid situation, the accuracy of pedestrian detection has a big challenge. The difficulties of the study lies in: looking for a more robust image characteristics description to describe sheltered pedestrians in complex scene, such features can not only describe pedestrian better, but also can distinguish the similarity of pedestrian effectively; finding a more effective detection algorithm of human body part model,which can efficiently identify pedestrians in complex scene. Aiming to solve this problem,this paper based on traditional pedestrian detection method of HOG + SVM emphatically discusses how to establish a more effective method to describe characteristics of pedestrians, make up the insufficient ability of expression for the image underlying characteristics and build a more effective part model for shaded pedestrian detection. From the perspective of the selection of characteristic expressions and human body part model, this paper conducts a thorough research on the pedestrian detection in complex scenes.Combined with the advantages of fisher vector, this paper proposes an algorithm of human body part detection based on fisher vector, to solve the problem of insufficient expression ability and low accuracy of detection when shaded pedestrian in conventional pedestrian detection. This algorithm uses fisher vector to quantify HOG feature of human body parts and full body, and obtains human body classifier and integral classifier through the training of support vector machine(SVM). Then Hough vote is used to vote for integral classifier and human body classifier,and thehighest score represents the pedestrians position. In the end, Non-Maximum Suppression is used to eliminate false alarm. Through the testing in the standard pedestrian database and comparing to the current mainstream pedestrian detection algorithm, in the end,INRIA pedestrian data is used to verify this method.In complex scenarios, some feature in a large number of overall feature are easily disturbed by environment factors and could not well represent the characteristics of shaded pedestrian. This paper put forward a kind of sparse feature selection algorithms to choose the most representative of pedestrian description characteristic in HOG feature. Then gradient descent optimizing method is adopted to sort the weight of DPM parts, and finally high accuracy classifier is formed by combining LSVM training algorithm.Finally,the experiments is designed for validation.This paper USES the MATLAB2014 B to complete the design of experiment and use the standard pedestrian database of INRIA to the contrast experiment, through the experimental results show that the proposed two methods of this paper has high accuracy and robustness in complex scene pedestrian detection.