Vehicle recognition in two non-overlapping views

The vehicles appearing in two non-overlapped views are different in size, aspect, color, and illumination. We apply the so-called non-metric distance embedding technique to identify the same vehicles. First, we extract edge and color features of the vehicles in the images. Second, we compute the difference between two vehicles and construct the distance feature matrix which can be used to select a set of example pairs. Third, we embed a pair of vehicle onto the example pairs to obtain the embedding feature vector. The images of the same vehicle in two views are embedded as positive-pair vectors, otherwise as negative-pair vectors which are used to train the SVM classifier. Finally, we show three different scenes for verifying the accuracy of the proposed method.