Research On Pedestian Detection And Tracking Technology In Indoor Scene

With the rapid development of computer vision technology,pedestrian detection and tracking technology has been widely used in robotics,traffic safety,video surveillance and other fields.In the background of complex reality,it will inevitably be impacted by the light and shades,pedestrian shelter,pedestrian attitude and other factors.How to accurately detect pedestrians and achieve rapid tracking is the core of research.Based on the analysis of pedestrian detection and tracking technology and the research status of related technology,this paper mainly uses Nao robot to collect video image data in the laboratory environment,and improve and verify the algorithm on the computer,and apply the improved algorithm to the Nao robot platform,so as to complete pedestrian detection and tracking experiments.Moreover and meet the friendly human-computer interaction.In order to achieve this goal,the research work of this paper is as below:In respect of pedestrian detection.Firstly,the pedestrian detection method of HOG feature algorithm is introduced in detail.Secondly,the HOG feature of the sample is extracted by using the INRIA pedestrian dataset and the image dataset in the laboratory environment,then the HOG feature of the sample is extracted to generate the HOG eigenvector.The extracted image dataset from the HOG feature is put into the SVM classifier for training.Finally,the image dataset taken by the Nao robot is used as the testing sample.Successful pedestrian detection in the image.In additional,based on the YOLO network model structure of depth learning,an improved pedestrian detection method is proposed based on the YOLO algorithm.The main improve method is to add the loss function of the YOLO network model.The pedestrian pictures are collected by Nao robot.The collected pictures are divided into S × S network cells,and pedestrians,who are falling into the network cells,are detected and predicted,meanwhile,the confidence score for each class in the border is calculated.Taking into account the loss of the detection of the border,The height and width of the border are replaced by their square root values.For the pedestrian image which falls into the border,their loss function are improved,so that the detection effect is enhanced further.Finally,the improved YOLO detection algorithm and HOG feature algorithm are compared in the same experiment conditions.In respect of pedestrian tracking.Based on the MeanShift pedestrian tracking algorithm,the Nao robot is used to collect the video image data in the absence of the obstacle laboratory environment,and the tracked pedestrians are selected.The MeanShift algorithm is iterated according to the histogram information in order to achieve the purpose of tracking.But there are problems with slow run speed of the algorithm and the search box can not adapt to the camera and track the target distance changes.So using improved Camshift algorithm based on the Kalman filter algorithm to search the size of window adptively,and then preditict the specific location for the next pedestrian.Thus solving the lost problem for pedestrian tracking when there exsits obstacle.But there is also the problem of slow run speed of the algorithm.Using KCF pedestrian tracking algorithm,in the presence of obstacles,the tracking experiments are completed for single or multi-person,so as to improve the accuracy and speed of the pedestrian tracking.From the accuracy and speed of algorithm,the pedestrian tracking algorithm is analyzed by comparative experiment.Under the Nao robot platform,an improved YOLO model detection algorithm is used to detect pedestrians,and KCF pedestrian tracking algorithm is used to realize the physical experiment of tracking pedestrians by the Nao robot.