Vehicle Detection Based On Spatiotemporal Feature

Compared with the traditional ground monitoring video,UAV aerial video has the advantages of large monitoring range,flexible,and can better detect the fast-moving target.For vehicle detection,aerial video coverage is complex,large area and small key target.At the same time,vehicle image is easily affected by illumination change,vehicle occlusion,camera jitter and other factors.Therefore,vehicle detection in the aerial video is still a challenging topic.This thesis mainly studies the detection of moving vehicles in aerial video.Aiming at the problem of low vehicle detection accuracy in existing algorithms,a vehicle detection algorithm based on spatiotemporal feature in aerial video is proposed.This thesis are divided into three parts: image registration,moving vehicle candidate area extraction and moving vehicle verification.In the image registration,SURF-based feature point matching method is used for background compensation.Because the traditional RANSAC algorithm is not effective in eliminating the wrong matching point pairs,this thesis proposes an algorithm based on the angle judgment to eliminate the outliers.The angles between all pairs of feature points are sorted in ascending order.The average value of the first 1/4 of all angles is used as the threshold,and the feature point pair whose angle and threshold difference is rounded down to greater than 0 is eliminated and the accuracy of the image matching is improved.The stage of aerial vehicle detection is divided into candidate regions extraction?rough detection?and moving vehicle verification?fine detection?.The rough detection is based on time feature,and uses the three-frame difference algorithm to obtain candidate regions that may have a moving target in the aerial video.Because of the camera's shooting angle and attitude change,target vehicle has a more complex form in the candidate area.The TBGC feature is proposed that increases the effect of the inverse clockwise neighborhood pixels and central pixels based on the BGC features of space.We extracted the TBGC features of the candidate regions and verified with a trained Support Vector Machine?SVM?to achieve fine detection and complete the aerial video vehicles detection.This algorithm is implemented on the MATLAB 2014 b platform.The experimental database adopts Munich Crossroad01 of VIVID's Egtest01-Egtest05 and KIT AIS,and two self-built data sets DJI₀006 and Video2.Experimental results prove that the TBGC features are more effective than BGC1,LBP,and HOG in vehicle detection,with an average detection rate of 93.09%.The TBGC features and dynamic features are combined to detect the moving vehicle in aerial video.The average F value reaches 93.73%,and the experimental results are better than a moving object detection algorithms for aerial video based on multi-model estimation.