Fine-Grained Vehicle Recognition via Detection-Classification-Tracking in Surveillance Video

Object recognition is a wide applied task in computer vision. Many fine-grained object recognition approaches are proposed in recent years to detect the same species objects effectively at subordinate-level. In this paper, we present a novel fine-grained vehicle recognition by utilizing collaborative feedback scheme of detection-classification-tracking in surveillance video. We collect a labeled data set of 200 car images which contains three types of car images in addition to negative samples, and extract Haar-like features from the training samples to build global appearance model for fine-grained feature representation. Then a multi-class SVMs classifier is trained offline on the training data set to distinguish the intra-class variability of cars. The collaborative feedback scheme incorporate the tracking and feedback constraint for reduce the frequency of detection and recognition. That is, the detector localizes the motion objects in every frame except that have been observed and tracked, and the best matching object with the given initial car is identified by the classifier and be tracked in the subsequent frames until it is not present. The collaborative scheme of detecting-and-tracking can decreases the computational cost in terms of the frequency of detection and recognition at each frame. The experiment result shows that our approach can effectively locate the best matching car at the frames when the target car appears.