Vision-based Scene Perception And Target Tracking For Unmanned Aerial Vehicles

In this paper,we study the problem of scene perception and target tracking using an UAV(unmanned aerial vehicle)with pattern recognition and computer vision skills.The mission of UAV scene sensing is to detect and identify important targets in the image of the drone’s camera and to accurately locate these targets.For the target of interest,control the drone to actively perceive and approach the target for stable tracking.This has important implications for security and daily life.For different scenarios and tasks,this article uses different methods to perceive the targets in the scene.For common indoor and outdoor scenes,this article trains the deep learning model based on Yo LOV3(You Only Look Once)to detect and identify 20 categories of common scene targets,including people,animals,vehicles and some indoor and outdoor items.For meaningful perceptual goals,such as pedestrians and faces,this paper compares the perceived effects of different methods on pedestrians and uses MTCNN(Multi-task Cascaded Convolutional Neural Networks)for face detection and key point location.For some special task requirements,this paper uses the image processing method in the virtual scene to effectively perceive.In the aspect of active sensing and tracking of special targets,this paper selects pedestrians as the tracking object,and the target pedestrian’s face as the target of active perception.Aiming at the mobile platform of UAV,this paper designs a method to solve the specific pedestrian tracking problem under multi-person interference,including two modules: feature extraction and dual queue target matching.In the case of smooth tracking of target pedestrians,this paper designs a positive face evaluation criterion based on the key points of the face,and according to the criterion,controls the drone to actively perceive and approach the positive face of the target pedestrian.Finally,this paper expands on the obstacle avoidance problem of UAVs,and proposes a monocular UAV obstacle avoidance method based on feature point method.In the experiment,a small drone with only a front monocular camera was used.Without GPS information,only the visual information was used to design a PID control algorithm to control the drone for sensing and tracking tasks.In this paper,a lot of experiments are carried out in data sets and outdoor scenes to verify the ability of drones to perceive and target tracking in actual scenes.In the aspect of obstacle avoidance using feature points,sufficient experiments were also carried out,and the task of crossing the woods was basically completed.