Improving Traffic Object Detection With Vehicle-Borne Fisheye Imaging

The image captured by the fisheye camera has a wide field of view, so, compared with an ordinary camera, the deployment cost is lower for sensing the same scene. However, due to nonlinear optical distortion, the object detection effect of the algorithm based on fisheye images is relatively poor. Therefore, the existing object detection algorithm cannot achieve a good detection effect on both fisheye and conventional images. Aiming at the problem that the detection accuracy and reasoning speed of existing object detection algorithms can not be balanced on fisheye images and conventional images, a general object detection model, fisheye road scene perception (FRSP-YOLO), which is based on the network model of you only look once version 7s (YOLOv7s). Firstly, the multi-branch attention weighting (MBAW)-efficient layer aggregation network (ELAN) module and the multiple revolution branches (MCB)-SPPCSPC module were designed to enhance and fuse multi-scale features. Then, a loss function was proposed to accelerate the model’s convergence speed and improve detection accuracy. Finally, the YOLOv7s network was improved by proposing a cross-level fusion mechanism, and the detection head was redesigned to improve object localization accuracy. The experimental results show that the FRSP-YOLO algorithm achieves high precision. The reasoning speed reaches 137.6 frames per second (FPS), meeting the requirements for real-time detection and laying the foundation for real-time autonomous driving. Its comprehensive performance is better than that of state-of-the-art (SOTA) object detection methods, enabling more efficient traffic object detection in complex road scenes.