Abstract: To address the problems of low detection accuracy and insufficient robustness caused by variable target scales and complex environments in autonomous driving, an improved depth and geometry aware object detection algorithm based on YOLO26, DGA-YOLO (depth and geometry aware YOLO), is proposed. Firstly, an R-G-Depth early fusion and iterative depth completion strategy is designed to enhance 3D spatial perception. Secondly, a Semantic-Geometric Aggregator (SGA) is introduced into the backbone to extract local semantics and macroscopic geometric context in parallel through two streams. Thirdly, an Anisotropic Geometric Convolution (AGC) is constructed, which parallels depthwise separable and pointwise convolutions to extract anisotropic geometric features, combined with a capacity-bounded scale-aware attention. Finally, a Large Selective Reconstructor (LSR) is deployed in the fusion layer, smoothing noise with large-kernel dilated convolutions and purifying geometric features with cross-strip convolutions. Experiments show that the proposed algorithm improves the mean Average Precision (mAP) on the KITTI dataset, with mAP@50 and mAP@50∶95 exceeding the baseline by 2.5 and 4.4 percentage points, respectively; it also significantly outperforms the baseline under physically simulated rain, fog, and rain-fog extreme conditions and in zero-shot generalization tests, demonstrating excellent anti-interference robustness and generalization capability.

Key words: autonomous driving, object detection, YOLO26, multi-modal fusion, depth and geometry awareness