Vehicle recognition is a key issue of intelligent transportation systems. For traditional color-based vehicle recognition methods, an important problem is that color information of vehicle is susceptible on lumination variation, which will inevitably reduce the recognition accuracy. Another is that there are contrastive interfering vehicles with similar color in same exposure environment, which could result in misrecognition of target vehicle. This paper presents a novel technique for recognizing target vehicle in different lumination scenarios via spectral feature. This method consists of three parts, ambient light spectrum calibration, vehicle spectral reflectivity detection and vehicle recognition. And an prototype system named Selectable Imaging Spectrum Detection System (SISDS), has been setup by this spectral feature-based method. A new combination algorithm is proposed for the detection part which contains the deep learning model YOLOv3 and the discriminant tracking algorithm KCF. Also, a CNN model is designed for improving the recognition accuracy through transforming the spectral data into two-dimension matrix. The experimental results of the SISDS clearly demonstrate that the spectral feature-based method overcomes the shortcomings of the traditional color-based method, and it can not only distinguish the vehicles of different colors in the case of overexposure, but also can recognize the target vehicle from the same color interference vehicles under various lumination, the recognition accuracy is up to 96.4%. Compared with color based method, the spectral feature-based method has great lamination robustness, and high recognition accuracy.
|