Research On The Moving Object Tracking Of UAV Based On Object Feature

Unmanned aerial vehicles(UAVs)have played an important role in reconnaissance,strike and monitoring in military and civil fields in tracking moving targets on ground.This paper mainly investigates the image representation model,the tracking of single moving target and of multiple moving targets by UAV under different influencing factors.Aiming at the influence of shadow interference,texture deficiency and tracking drift on the target under the influence of environmental factors,the corresponding model of target feature is established.Aiming at the impact of the carrier aircraft and the camera movement like the jitter of the scene,the background rotation and the target deformation on the background and the target,create the corresponding background model or the target feature model.Aiming at the shadow interference,lacking of texture and tracking drift in the real-time tracking of small targets on the ground by UAV,an algorithm of tracking single moving target on the ground by UAV based on improved correlation filter and multi-feature is proposed.Extract multiple features of the target: HOG feature,Color names feature(CN),Convolutional neural network feature(CNN).The spatial reliability graph is used to improve the correlation filter to improve the tracking performance of irregular objects.Based on efficient convolution operator,feature dimension reduction and compact sample space are constructed to significantly reduce spatial and time complexity and provide better sample diversity.The position of the target tracking box is determined according to the target characteristic response obtained by the correlation filter,and the target tracking is accomplished.Experimental results show that the algorithm can track single small moving target accurately in real time,with an average frame rate of 44.09 frames/sec,and has preferable real-time performance.In order to solve the trajectory of discrete and large state space over time of multi-target tracking track dispersion,a UAV tracking algorithm based on the fusion of data association and trajectory estimation is proposed.The discrete data association and continuous trajectory estimation are fused into an energy model based on the label costs,and the optimal multi-motion target trajectory distribution is obtained by minimizing the energy function.Label costs are used to prevent alternative minimization from overfitting,and are used as a model option to standardize the number of tracks.The appearance model based on color feature is added to the label costs to improve the robustness of the algorithm.In order to solve the problem that multi-target is prone to happen identity exchanging and re-numbering in staggered scenes,a multi-target authentication algorithm integrating color histogram and color distribution is proposed to compare the similarity of each track target and accomplish the continuous tracking of multi-target.Experimental results show that this algorithm has a stable tracking effect for different multi-target tracking application scenarios.