Vehicle identification based on feature point matching and epipolar geometry constraint
Vehicle identification across non-overlapping cameras has wide applications such as parking management, speed estimation and vehicle tracking. In these applications license plate plays an important role, [13, 14, 15] use license plate characters to identify vehicles by exploiting image processing technology. However, when the vehicle does not have a license plate or the license plate is occluded, other features of the vehicle have to be employed. Shan et al.  propose a novel solution converting problem of identifying vehicles across non-overlapping cameras into a same-different classification problem without direct feature matching by computing the same-different probabilities. Ferencz et al.  propose an on-line algorithm building an efficient same-different classification cascade by predicating the most discriminative feature set for vehicles, not only estimates the saliency and scoring function for each candidate feature, but also models the dependency between features. Wang et al.  tackle
vehicle identification by reconstructing vehicles with multiple linear regression models and sparse coding which has been successfully used in multi-samples classification and identification. Wang et al.  introduced an inter period adjusting technique based on the exponential smoothing to define an appropriate time-window constraint to identify vehicles. Matching vehicles which are subjected to drastic pose change and extreme illumination variation is conducted in , Hou explicitly estimates pose and illumination effectively for vehicles in reference image and target image respectively using approximated 3D vehicle models and albedos estimation, then re-render vehicles in reference image according to vehicles’ pose and illumination in target image to generate relit image, finally comparisons is made between the relit image and the re-rendered target image to determinate whether the vehicle in the reference image is identical to the vehicle in target image. Tian et al.  use multiple sensors to accomplish vehicle identification; vehicle status and correct signature segmentation can be
determinated by the matching result of one vehicle’s signature obtained by different sensors. The co-relationship between signatures can be obtained, and then the time offset is corrected depends on such a co-relationship. Sanchez et al.  propose a method based on matching electromagnetic vehicle signatures, which are obtained from wireless magnetic sensors. Jazayeri et al.  apply HMM (Hidden Markov Model) to separate background and moving vehicles in the temporal domain. Their identification process is based on vehicle and background motions.