Research And Application Of Human Detection And Tracking Methods Based On Family Scenes

Human detection and tracking technologies are commonly used in intelligent security,intelligent space,intelligent elderly care,and home service robots in the home scene.Human body detection is the basis of other human image processing tasks.In family scenes,the non-rigid body of human body and the complexity of environment bring a series of difficulties for human body detection.Human tracking can be divided into two aspects: image human tracking and scene human tracking.In the aspect of image human body tracking,most scholars focus on improving the speed and accuracy of target tracking algorithm,but there are few researches on abnormal tracking.Scene human tracking can give the location of the human body in the scene.However,the human body tracking algorithm based on the main axis can't deal with the occlusion problem.However,another pixel-based human body tracking algorithm is not suitable for human body tracking in family scenes due to its poor real-time performance due to complex calculation.In the aspect of the human body detection,the latest research results in this paper,the target detection of YOLO V3 algorithm,analyses the algorithm of detection performance in the scene in the family of the advantages and disadvantages,homemade family scene human data set,get the anchor clustering,using Densenet network unit instead of part of the residual network unit,improve YOLO V3 network structure,has the stronger ability of feature extraction,and combining with moving objects algorithm and frame the way to improve the efficiency of detection.In terms of image human body tracking,KCF target tracking algorithm has strong advantages over other algorithms in terms of speed and accuracy.However,KCF has two obvious defects: first,model drift is easily caused by the fixed scale;Second,KCF updates the model without considering the accuracy of tracking results.Aiming at the first problem,this paper designs a scale estimation algorithm to solve the model drift problem based on the scene information.For the second problem,an abnormal tracking criterion is designed to determine whether the tracking results are normal or not,and then the tracking results are fed back into the model.Human body tracking for the scene,using detection or tracking rectangular box instead of human prospect information,using the single head should be projection method and the single foot should be projection method to generate the candidate point feet,according to certain rules of clustering,arrowheads candidate foot points within the distribution value according to the distance to the center of the cluster,according to the weight threshold delete some candidate foot points,finally obtain the human position in the scene.Related experiments and the simple family scene human detection and tracking system established show that the improved YOLO V3 algorithm has lower missed detection rate for human detection in the family scene,and it also has good detection performance for fuzzy targets,small human targets and covered human targets.The improved image human tracking algorithm effectively solves the problem of tracking failure caused by human occlusion,blurring and loss.At the same time,the scene human body tracking algorithm designed in this paper has been greatly improved in real-time and accuracy compared with other scene human body tracking algorithms.