Accurate Target Tracking And Scene Understanding For UAV Platform

In this paper,computer vision and machine learning methods are used to study target tracking under severe occlusion and UAV platform scene understanding.Target tracking plays a vital role in many visual applications,such as UAV tracking,motion analysis,and video detection.And the scene understanding based on unmanned aerial vehicle?UAV?platform also has a broad application prospect.The current target tracking algorithm is faced with many tests.When the tracking target is seriously blocked,the partial features of the target disappear,which poses a great challenge to the robustness of the algorithm.Therefore,this paper proposes a Gaussian joint model?GSCM?with sparse discriminative classifier?SDC?and sparse generation model?SGM?for target tracking under severe occlusion.Among them,in the SDC model,the candidate samples are weighted according to prior knowledge of Gaussian distribution,and the weights of candidate samples in the current frame are predicted according to the variance and mean of the previous frame.At the same time,both the confidence of samples and the similarity between samples and templates are calculated by using both l₁ norm and LLC simultaneously in SDC and SGM models,and the coefficients obtained from them are combined.Finally,the sample with maximum likelihood is decided by weight,confidence and similarity.The proposed method not only considers the sparse representation ability between samples and templates,but also shows the high accuracy of l₂ norm reconstruction and avoids the error accumulation caused by the too small l₁norm reconstruction error,which causes the target to lose and track the failure.In the context of understanding,this paper uses deep learning target detection method YOLO v2 for UAV platform,fine-tuning the network,training five real-time detection five things model.In this paper,we use the public OTB dataset to select 17 groups of video sequences with serious occlusion.We use 12 different trackers in recent years to compare the experiments,compared to the state-of-the-art algorithms,our scheme can achieve better effects in heavy occlusion video sequences for tracking target.To the scene understanding,the UAV with a single camera,and without the help of GPS signal,a lot of experiments are carried out in a real environment by using a self-trained deep learning model.Complete the scene understanding in the real environment and let the unmanned aerial vehicle find the door to fly into the door.