To address the challenges of binocular vision ranging under complex environmental conditions—such as illumination variations, occlusion, and textureless regions, which result in unreliable and non-robust performance—this paper proposes a multi-source heterogeneous sensor fusion ranging method integrating 4D millimeter-wave radar with the YOLOv5-Monster framework. This method is capable of overcoming the issue of limited ranging accuracy in monocular or binocular vision algorithms under non-ideal imaging conditions. This study achieves high-precision spatial perception through the following specific pipeline: First, Zhang’s calibration method is used to obtain the intrinsic and extrinsic parameters of the binocular camera, and stereo rectification is performed on the raw images. Next, a lightweight YOLOv5 network is employed for object detection, while a high-performance Monster network is utilized to generate dense disparity maps, thereby accomplishing initial depth estimation. To mitigate the inherent depth estimation errors of vision-only systems, 3D point cloud data from a 4D millimeter-wave radar is further introduced. By applying a Kalman filter algorithm, the millimeter-wave radar point cloud and visual outputs are fused, achieving spatiotemporal synchronization and optimal state estimation across modalities and effectively correcting biases in visual ranging.
Experimental results show that within the full range of 4 to 150 meters, the relative error of the proposed method remains below 5%. Specifically, the relative errors are 1.25% (absolute error: 0.05 m) at 4 meters, 1.40% at 5 meters, 2.99% at 75 meters, and 4.91% at 150 meters. Compared with the vision-only Monster-YOLOv5 baseline method, the relative error at 150 meters is reduced from 13.16% to 4.91%, representing an accuracy improvement of over 60%. Meanwhile, in terms of long-distance error control, the proposed method significantly outperforms traditional stereo matching approaches such as SGBM+YOLOv5 and BM+YOLOv5, reducing errors by more than 20 percentage points. These results verify that deep multi-modal fusion can enhance environmental adaptability and measurement reliability, providing a high-precision and highly robust solution for distance estimation in intelligent perception systems, which holds important theoretical and engineering significance.