Scale and Occlusion Aware Object Detection for Autonomous Driving

Accurate and real-time object detection is the most fundamental requirement in autonomous driving systems to make safe navigation decisions. Robust perception of small and occluded objects is one of the most challenging problems in this field. Existing methods often result in false negatives and misclassifications due to the key reasons: inherent spatial resolution constraints in feature representations, insufficient contextual reasoning capabilities, and sub-optimal multi-scale feature integration. To overcome these deficits, a new detection framework based on YOLOv5 is proposed, featuring two complementary attention-augmented modules: the Multi-Scale Context-aware Attention (MSCA) block and the Atrous Dual Pooling Attention (ADPA) block. MSCA in the deeper backbone layers enhances semantic context reasoning for occluded objects, and ADPA in the neck pathway enables fine-grained detail preservation for small objects. Adaptive Target Weighting Loss (ATWL), a new loss function, is introduced that dynamically emphasizes small and heavily occluded targets, improving recall in challenging scenarios with a well-balanced precision-recall trade-off. Collectively, these architectural innovations establish a strong balance among accuracy, robustness, and efficiency, thereby addressing the complex demands of autonomous vehicle perception under real-world conditions. Extensive tests using challenging benchmarks demonstrate the effectiveness of the proposed approach, with mAP of 87.7% on KITTI, 53.4% on BDD100K, and 37.6% on the Indian Driving Dataset (IDD), while maintaining computational efficiency and real-time processing.