Mixed-initiative Recognition And Tracking Of Ground Vehicles Using UAV Images

The perception and understanding of the UAV(Unmanned Aerial Vehicle)’s ground target is one of the typical tasks of the UAV system,and they have received extensive attention.At present,the UAV system has limited autonomy in complex environments.In particular,the target-aware task still needs people to participate in order to ensure the reliability of the task.Therefore,the research on the target-awareness method of humancomputer cooperation is of great significance in practical application.This thesis focuses on the key task of mixed initiative sensing,and studies the detection,identification and tracking methods of UAVs on ground vehicle targets,including initiative sensing of UAVs and initiative sensing of operators.Initiative sensing of UAVs is the use of machine vision algorithms to perceive the ability to capture target features at the data layer.Operators’ initiative perception means that operators are aware of the target at the task level based on cognitive mechanisms such as visual attention.They work together to accomplish the target-awareness task.In addition,due to the particularity of military missions,it is often difficult to obtain sufficient target sample data.The focus of this paper is under the background of small data and mainly on the following aspects:(1)For the problem that the accuracy of object detection of edgeBox algorithm is affected by background such as shadow,an SR-edge Box detection algorithm combining frequency significance detection and object detection is proposed.By suppressing the low frequency in the frequency domain,highlighting the high frequency,and enhancing the structural region in the time domain to weaken the irregular shape background,more accurate target detection without specific target prior information is achieved.The images of the shadowed vehicle taken by the UAV are tested to verify that the proposed algorithm has higher detection accuracy under the shadow condition than the original edge Box algorithm.At the same time,for the problem that the SR-edgeBox algorithm detects many regions and is not conducive to subsequent recognition,an automatic candidatebounding-boxes clustering algorithm is proposed.The candidate bounding boxes belonging to the same target are grouped into one class based on the overlapping.The clustering test results of various target scenarios verified the effectiveness of the proposed clustering algorithm.(2)For the traditional recognition methods based on corners,colors and other features in the UAV application,there exist some problems such as unstable features,occlusions,demanding much sample data,less of semantic interpretability and so on.This thesis proposed a recognition method of vehicle using key parts under small sample data.The method recognizes vehicles by segmenting and recognizing the parts of the car and then inferring under the Bayes framework.Among them,the existing iterative graph cutting algorithm(GrabCut)is difficult to effectively segment the deficiencies of the components,the maximum inter-class variance method(Ostu)is used to calculate the optimal segmentation threshold,and the iterative graph cutting algorithm based on random scribing is designed to realize fine segmentation of the components.Concerning about the problem that the traditional method is difficult to deal with the deformation component identification,a method by building the component probability map model is designed.The probability map model is established by sparsely representing the contours of the component using oriented edge nodes and constraining the continuity of the contours.Based on the prior observation information and the identified components,Bayesian is used to reasonably identify whether the target is a vehicle.The image test with UAV shows that the recognition method can identify the vehicle target with a high accuracy rate when the sample size is small and there exists occlusion,and the introduction of the component concept makes the recognition method semantically explainable.(3)Concerning about the problem that the automatic tracking algorithm is unstable in complex environment,taking TLD tracking algorithm as an example,this thesis proposes a robust tracking hTLD(human-in-the loop TLD)algorithm with human in the loop.The algorithm comprehensively utilizes human cognition ability and automatic algorithm’s ability of capturing the target by designing an online interactive learning mechanism.The human is responsible for the task-oriented tracking of the target from top to bottom,and can introduce the cognitive information into the algorithm;the automatic algorithm is responsible for the bottom-up data-driven tracking of the target and provides the tracking results and confidence to the human,what’s more,at the same time the human cognition information is incorporated into the algorithm for online learning.Offline verification on the video taken by UAVs shows that the tracking speed and accuracy of the hTLD algorithm is improved compared to the original TLD algorithm.And with the interaction with human,the false alarm rate and the missed alarm rate of the confidence alarming system are reduced,and the human can rely more on the intuitive confidence prompt,so that the work load can be gradually lighten.