Research Of Moving Target Detection And Tracking Algorithm Based On Machine Learning

Unmanned aerial vehicle(UAV)has attracted wide attention in various fields due to its small size,low cost,flexibility and easy operation.It is in great demand in the fields of emergency rescue,agricultural plant protection,military reconnaissance,geological exploration,environmental monitoring,film and television entertainment,etc.Visual object detecting and tracking is one of the fundamental contents in Drone aerial photography,In order to realize the tracking and shooting of the UAV,the detection and tracking of moving targets is essential.This paper adopts the random forest and KCF algorithm,focusing on the basic subject of moving target detection and tracking,studying the feature extraction,boundary detection algorithm and occlusion detection mechanism of moving target,and finally transplanting the EBKCF algorithm to UAV to verify the algorithm.Firstly,a simple gradient extraction algorithm Simple-gradient is therefore proposed to solve the problem that the feature extraction algorithm of structured random forest is complex and the detecting time of moving object is too long.The algorithm reduces the complexity of feature extraction,shortens the extraction time of the algorithm,the training time of the moving object model and the detection time of the moving object boundary.Aiming at the redundancy among multiple decision trees,the pruning operation is made to speed up the detection speed of moving targets.Secondly,aiming at the moving target tracking algorithm of Kernelized Correlation Filters(KCF),an occlusion detection mechanism MPCE index is proposed to solve the problem of moving target tracking caused by occlusion in the visual field.Whether there is an occlusion in the current line of sight according to the relationship between the occlusion detection factor and the threshold,to improve the tracking accuracy of the algorithm under occlusion,Aiming at the drift of tracking target caused by the scale change of moving target,the size of moving target is updated by Edges Boxes algorithm.At the same time,a proposal veto mechanism is utilized to obtain more accurate moving target frames for multiple candidate results.According to the experimental verification,this algorithm is a real good performance in the face of occlusion and scale changes.Finally,in order to verify the effectiveness of EBKCF algorithm,the guidance vision system combined with Manifold airborne computer is used to carry out moving target tracking experiment on the DJI M100 Unmanned flight platform.The EBKCF and the KCF algorithms are transplanted into Manifold,and according to the current experimental conditions designed the experimental procedure.The comparison experiments between the KCF algorithm and the EBKCF algorithm show that the latter algorithm can accurately track the scale changes and occlusions of moving targets.