Study On Vision-based Vehicle Proposal And Vehicle Detection

Vehicle detection based on vision in unmanned and Advanced Driver Assistant System draws widely attention and makes extensive application.Most of traditional detection algorithm based on sliding windows for test,and the use of proposal can effectively reduce the amount of detection window,and speed up the efficiency of detection.In this paper,based on vision,on road front-vehicle detection is studied,the main work and novelty includes:1.In this paper,we provide a novel approach to get front-vehicle proposal for vehicle detection based on accurate structured edge detection.The algorithm uses structured random forests to detect edge of the vehicle accurately.The edge of the vehicle incorporates more labeled information,and integrate the vehicle overall and local interests.Directly through the edge candidate box,overcome the problem of the vehicle scale selection,get accurate vehicle candidate box.After projecting edge on the horizontal or vertical and extracting peaks,we can get the candidate edge.With the help of integral image,we can find the proposals quickly,linear SVM and 4D feature is adopted to rank and restrain the proposals.Using Non-Maximal Suppression(NMS),we can balance the numbers of the proposals and the precision.Space scale constraint is used to further remove unlikely candidate window,to reduce the number of proposals.2.We compare our method with BING,Edgebox,Selective Search which are all state-of-the-art proposal methods in KITTI database.After evaluation,it shows that our method has a good performance on the number of proposals,the detection rate of proposals and the precision of proposals.Two methods are proposed to combine proposals and the after detector algorithm.One is to resize the proposals size,the other is to generate candidate area.We combined our method with ACF detector and test on KITTI database.Finally,the results show that the combination reduce the number of detection windows and the False Positive Rate.3.Based on Zynq-7000 Extensible Processing Platform,we realize the function of vehicle detection.we also implement the image acquisition and label the vehicle on road.Finally,we preliminarily validate of the function of vehicle detection using the true data..