基于单目视觉的夜间车辆识别方法

Vehicle identification method in nighttime based on monocular vision

快速准确地识别车辆是夜间行车安全预警系统的关键技术。提出在RGB空间下基于夜间尾灯图像的实时车辆识别方法。通过对夜间尾灯颜色特征进行分析和提取,选用尾灯图像的(R-G)色差特征作为图像分割的输入,并采用自适应阈值法分割图像。借鉴非最大抑制算法的思想,通过设置尾灯粘连区域判断条件,解决夜间尾灯粘连问题;根据同车左右尾灯区域特点,定义尾灯配对规则和尾灯差异度,实现尾灯配对;并定义车辆包围框和设置最大重叠率的方法,解决夜间车辆重叠问题。实验结果表明,该方法计算量小,平均每帧耗时26 ms,在含有尾灯粘连和车辆重叠混合场景下,车辆识别准确率大于93%。

Fast and accurate vehicle identification is the key technology of safety early warning system for driving at night. The paper proposes a real-time vehicle identification method based on image of night taillights in the RGB color space. For analyzing and extracting color characteristics of taillights, the chromatic aberration characteristics of（R-G） image are selected as the input of image segmentation, for which the adaptive threshold algorithm is used. Solving the problem of taillights adhesions at nighttime, the adjoined taillights are cut apart by setting the judgment condition of adjoined taillights, which is inspired by the non-maximum suppression algorithm. According to the similar characteristics of the left and right taillight on a car, the taillights are paired by defining the rules of taillights pairing and the difference degree of taillights. And the overlapping image of vehicle is handled by defining vehicle bounding box and setting the maximum overlapping rate. The experimental results show that the method has small amount of calculation, average time-consuming per frame is about 26 ms, and the accuracy rate of vehicle identification is more than 93%under a mixed scenes containing taillights adhesion and overlapped vehicles.