Vision-based Pedestrian Tracking And Scene Understanding For Unmanned Aerial Vehicles

In this paper,we study the problems of pedestrian tracking and scene understanding for unmanned aerial vehicles(UAVs)with computer vision and pattern recognition skills,and the study of pedestrian tracking is the main part.Pedestrian tracking for UAVs is important for ensuring public safety,and scene understanding has broad prospects for development.In terms of pedestrian tracking,we present a new method in order to solve the problem of the tracking of a specific target pedestrian under multi-disturbance conditions,and the method is consisted of pedestrian detection,feature extraction and target matching.In the pedestrian detection part,we adopt the classical method of HOG + SVM,and we use blocked image color feature in the feature extraction part.In the targetmatching part,we introduce a queue to store the features of the target pedestrian,and use LLC to reconstruct any pedestrian feature based on the feature queue,and then make decisions according to the reconstruction error and maintain the feature queue with certain rules.In terms of scene understanding,we apply YOLO,which is one of the deep learning object detection methods,on UAV platforms.We adjusted some network parameters of YOLO and trained a model that is capable of detecting and recognizing pedestrians,cars,buildings,trees and doors in real time.Each UAV we adopt in experiments is only equipped with one front camera,and the motion control of the UAVs is based on the classical PID principles,without any GPS information.In terms of pedestrian tracking,we compare the tracking method we present with some well-known object tracking methods in terms of accuracy and real-time performance based on test datasets we collected and one public test dataset downloaded from the Internet,and the experimental results verify the effectiveness of our method.In terms of scene understanding,we also conduct numerous experiments in outdoor conditions with the model we trained,and realized the goal of letting the UAV fly into the building through the glass door piloted by itself.