Vehicle Detection Via Saliency Detection And Classifier Training On Aerial Image

Vehicle detection in aerial images, which is importantly applied in military, geographical information and other fields, is a significant part of intelligent transportation. Target detection technology, which is core technology involved in this field, is a classical problem in computer vision and machine learning. Compared with ordinary natural images, aerial image has the following characteristics:low resolution, insufficient details, photos focus hard; target is too small, hard to detect; road ambient occlusion, shadow, texture of the background interference greatly. These particularities make it impossible to achieve the acceptable effect by applying the classical method of target detection in aerial images. As a resolution to this problem, this paper proposes a two-stage method for vehicle detection in aerial image, the first stage is road detection, phase Ⅱ is vehicle detection based on training classifiers. The two stages cooperation makes it possible to achieve robust vehicle detection in aerial images. Using automated assessment techniques to evaluate the aerial vehicle detection and experimental results verify the accuracy of the method in this paper.The main contributions of this thesis are as follows:(1) For road detection in first stage, this paper proposed road extraction based on histogram contrast and post-processing methods. Method based on histogram contrast is able to obtain most road region, but remains problem in integrity and disruptors removing, therefore, we use the combination of thresholding and morphological approach to solve the remain problem. At this stage, we effectively extracts the road from images, and the results are compared with other methods on both efficiency and evaluating indexes, experimentally certificated that road extraction algorithm in this paper can improve the accuracy of subsequent vehicle detection.(2) For the second-stage, vehicle detection based on classifier, in training, we use the combination of Haar features of vehicle and Adaboost classifier for training. In tests, based on the road detection result of the first phase, using a sliding window search and classification identification technology for detecting vehicles. Through extensive experiment researching on choice of classifier, the size of the training set and the search step size selection influences on performance of vehicle detection. This article also uses an evaluation approach to automatically evaluate vehicle detection performance. Experiments showed that set the tested-error sample back to training can greatly reduce the error rate; choose the right search step can achieve the optimum balance between accuracy and efficiency.GroundTruth sets, which is marked in this theses can be applied in aerial vehicle research in image-related areas, the relevant laws concluded from classifier detection technology are useful for aerial vehicle detection and other object detectionfields.