考虑广角镜头畸变的精确目标识别距离估算

人脸识别和车牌识别是智慧安防领域的重要内容。人脸和车牌的特征尺度小、细节丰富，对成像系统的空间分辨力有较高要求，需要较大规模的探测器和高传函、小畸变的光学镜头。然而，安防系统又要求广域监控，需要使用视场大但具有一定畸变的广角镜头。因此，设计既能精确识别人脸和车牌目标、又能广域监控的成像系统时，应将精确目标识别作为约束来权衡高空间分辨力和大视场的性能指标以及估计识别距离。在这样的应用需求下，本文提出了像素面密度对精确目标进行统一描述，并提出了考虑广角镜头径向畸变的精确目标识别距离估算方法，通过对存在旋转和平移的人脸和车牌目标进行计算验证，结果表明:考虑径向畸变后实际识别距离较理论识别距离近，且人脸和车牌平移距离分别为1 m和2 m时，实际与理论的识别距离差异高达34.2%和27.5%。

Distance Estimation for Precise Object Recognition Considering Geometric Distortion of Wide-angle Lens

Face and license plate recognition are crucial aspects in the field of intelligent security. A high-spatial-resolution imaging system with a large-format detector and low-distortion optical lens is required for recognizing small-scale features and rich details in faces and license plates. However, security systems need to monitor wide area, which requires a wide-angle lens with a wide field of view, but with some distortion. Therefore, precise target recognition should be used as a constraint to balance the high spatial resolution and wide field of view when designing an imaging system that can recognize details and monitor a wide area. Under such application requirements, an evaluation index based on pixel areal density is proposed. With the aid of this evaluation index, a distance estimation method for precise object recognition, considering the radial distortion of the wide-angle lens, was designed. Rotated and translated faces and license plates were used to demonstrate the estimation method. The results indicate that the recognition distance with radial distortion is less than that without radial distortion. When the translation distance is 1 m and 2 m, the difference between the actual recognition distance and the ideal recognition distance is 34.2% and 27.5%, respectively.