Research Of Pedestrian Detection And Tracking Algorithm Based On Learning

With the development of intelligent life concept,the concepts of intelligent video surveillance,intelligent assistant driving and home automation have been widely concerned.As the core technology,the pedestrian detection and tracking algorithm has received more and more attention of the researchers.This thesis aims to achieve a robust pedestrian detection and tracking algorithm based on learning.Firstly it detected the two parts of human which have features of face or body,then started tracking the pedestrian with automatical initialization.This thesis focuses on the detection algorithm based on Adaboost and tracking algorithm based on Online Boosting.According to the existing problems,we improved the detection and tracking process respectively.The main research contents in this thesis are as follows:In the aspect of detection,this thesis realized Adaboost face detection based on Haar features and Adaboost body detection base on HOG features firstly.Then we changed the feature to LBP feature to solve the problems of the long detection time and the high missing rate owing to the rotating target and changing illumination in the detection.In the process of LBP feature extraction,we combined the LBP feature based on rectangle with the rotational invariance LBP operator.The improved LBP feature is more descriptive.The miss rate of face detection results decreased from 7.6%to 5.5%.The false positive rate of body detection results decreased from 11.5%to 8.9%.At the same time,the time of the detection was greatly reduced.In the aspect of tracking,we replaced the Haar features by LBP features.The weight of the original classifier is also changed when the classifier is updated.Then the optimal weight of each classifier is gained.At the same time,we combined particle filter and Online Boostingalgorithm.We calculated the similarity between the targets' features detected by the Online Boostingstrong classifier and the features of each particle to calculate the weight of each particle,the tracking drift phenomenon was improved.Also the tracking was more accurate.Compared with the traditional Online Boostingalgorithm,the average center deviation of the improved algorithm decreased from 13.6 to 7.9.The time for the tracking Shortened from 588ms to 183ms.Based on the above research contents,this thesis implemented a complete automatic pedestrian detection and tracking system.In the nearby scene,The AdaBoost algorithm based on Haar and LBP feature is chosen to detect face,And in the faraway scene,the AdaBoost algorithm based on HOG and LBP features can be used to detect body.Moreover,the system includes tracking algorithms based on online Boostingand online Boostingembedded particle filter.The system can automatically detect the target and start tracking after the video is loaded.Experiments proved that the system could detect the pedestrian in the image accurately and carry on the tracking stably.