Research On Important Issues Of Pedestrian Detection And Appearance Feature Analysis

Pedestrian detection is a hard and hot topic in the field of target detection, and has been widely used in the intelligent video surveillance system, video analysis system, intelligent transportation system and driverless system. The varied appearance and non-rigid shape of the pedestrians increase the difficulty of effective pedestrian detection. As the application scenario has become more complex, demand facial features, clothing colors and more information of pedestrians in the system such as criminal investigation and video security. This information provides sufficient basis for the future study of pedestrian tracking and pedestrian recognition in different areas. This paper mainly study on pedestrian detection and appearance feature analysis. This paper studies and realizes three parts:pedestrian detection, pedestrian’s facial feature analysis and pedestrian’s clothing feature analysis, mainly includes these following works:(1) Research and implement pedestrian detection based on deformable part models. DPM use improved HOG feature as pedestrian’s feature descriptors. The improved HOG feature has less dimensionality and more robust to illumination and shape variations than original HOG feature. DPM use Latent SVM learning the improved HOG feature of the whole pedestrian target and the parts of pedestrian target, build the human star-structured model which can be used to detect pedestrian and the parts of pedestrian. Based on the human star-structured model, algorithm can locate the head area, trunk area and leg area. Lay the foundation of follow-up studies for pedestrian’s facial feature analysis and pedestrian’s clothing feature analysis.(2) Propose a method for pedestrian’s facial feature analysis, includes face detection and feature points locating based on the human star-structured model. The head model of human star-structured model can locate the pedestrian’s head area. Use Haar feature and Adaboost cascaded classifier to detect face on the head area. Extract local binary features from face based on random forest. Learn global linear regression model from all the local binary features and locate the facial feature points. The method for face detection and feature points locating based on human star-structured model reduce the time from scale and iterate through the images. Reduce the time of face detecting and feature points locating.(3) Propose a method for pedestrian’s clothing feature analysis, includes clothing area segmentaion and color recognition based on the human star-structured model. Combine some specific part models as trunk model and leg model. Use GrabCut to segment the pedestrian’s clothing foreground area from the area which is located by trunk model and leg model. Reduce the impact to clothing color recognition caused by background area. GrabCut using Gaussian mixture model describe the pixels’variance and mean from clothing foreground area in RGB color space. According to the pixels’ mean and quantity implement pedestrian’s clothing color recognition in HSV color space.This paper use human star-structured model based on DPM to detect and recognize pedestrian’s facial feature such as face, eyes, nose and mouth, clothing feature such as clothing color. This paper use pedestrian detection, face detection, facial feature points locating, image segmentaion and color recognition to describe and analyze pedestrian comprehensively and systematically.At last, this paper introduced the pedestrian detection and appearance analysis system. This system is developed by C++ and Opencv, using the methods researched in paper. The framework and UI of system is designed by Qt. Display the result of detection and analysis vividly. This system can set the abnormal threshold number of detecting and analyzing results including pedestrians, faces, facial feature points and certain color of pedestrians’clothing, save the result and video frame to MySQL database.