Research On Video Processing Based Specific Target Detection Technology In Mini UAV System

Target detection and tracking in mini UAV video have important research values and wide applications in sercuity,terrorist defence and military affairs.Regard pedestrian as specific target in UAV video,pedestrian detection in various sences is studied by visual saliency and machine learning.Pedestrian tracking is optimized by mean shifting.This specific pedestrian identity is finally studied via human facial character description so as to achive the specific target detection under mini UAV aeial photo system.The main work of this paper includes the following aspects:1)For pedestrian detection in UAV video,a two-stage pedestrian detection method based on Graph-based Visual Saliency(GBVS)and Aggregate Channel Features(ACF)is proposed.First,orientation and color features of GBVS model are modified for pedestrian detection.Also,the final saliency map is formed by weighted merging instead of a direct sum of feature channels.Pedestrian candidate region is obtained by saliency detection,realizing the rough detection of first stage,also narrowing the image region needed to be processed in the second stage.Considering ACF algorithm's advantage of fast multi-scale detection,in the second stage,the pedestrian candidate region is further analyzed by ACF algorithm to determine whether it contains pedestrian or not.LUV,gradient magnitude and gradient histogram are extracted as pedestrian features,and then trained by AdaBoost algorithm to construct a strong classifier,finally fast feature pyramids method is used for fast multi-scale detection.Experiment results show that our method achieves higher detection rate and less false positives than ACF algorithm.2)For pedestrian tracking in UAV video,a modified mean shift algorithm based on multi-feature fusion and back projection is proposed.Aiming at the tracking instability caused by single feature in traditional mean shift algorithm,color and texture feature are fused for the description of the target.Background weighting technique is adopted to enhance the accuracy and robustness of target feature model.In order to overcome the tracking failure caused by the size-fixed tracking window of the traditional mean shift algorithm,the trend of target size is first determined by back projection,and then a further judgment is made by similarity comparison to decide whether the tracking window size needs to be adjusted.The PETS2009 database and practical video were selected to verify the effectiveness of this tracking algorithm.3)For pedestrian recognition in UAV video,face detection and recognition technology based on frontal face is applied to realize pedestrian recognition.First,rough face detection is realized by AdaBoost classification algorithm based on haar-like feature.Then,a Gaussian skin model based on H-CgCr region segmentation is proposed.Skin detection is performed on the rough detection windows to reduce the occurrence of false positives.After face detection,Gabor filter bank is constructed to extract the Gabor feature of human face to enhance the robustness to environmental illumination change.Then,the principal feature space is obtained by two-dimensional principal component analysis(2DPCA)method.Finally,face recognition is realized by similarity comparison between the projection coefficients of input face and training samples in the principal feature space.