HOPNet: Heterogeneous Object Priority Network for Unusual Object Detection in Mixed Traffic

Engineering notes

Key topics: autonomous driving, autonomous vehicle, object detection, lidar, camera-lidar fusion, perception. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Indian roads present one of the world's most heterogeneous and unstructured traffic environments, simultaneously accommodating motorized vehicles, e-rickshaws, bicycles, pedestrians, and free-roaming animals without strict lane discipline. Autonomous vehicle (AV) perception systems trained primarily on Western datasets fail catastrophically in such mixed-traffic scenarios. This paper proposes HOPNet (Heterogeneous Object Priority Network), a multi-modal real-time object detection framework integrating YOLOv8-based convolutional neural networks with Swin Transformer attention, camera-LiDAR fusion, and domain-adaptive transfer learning. HOPNet is trained and evaluated on the Indian Traffic Dataset (ITD), a novel dataset of 180,000+ annotated images collected from twelve Indian cities. Experimental results demonstrate a mean Average Precision (mAP@0.5) of 91.4%, surpassing the best baseline (YOLOv7) by 11.1%, with a detection latency of 28 ms on NVIDIA Jetson AGX Xavier edge hardware. Per-class analysis shows particularly significant improvements for underrepresented categories: monkey (+27.6%), goat/sheep (+26.2%), street dog (+21.9%), and loaded bicycle (+18.5%). The proposed Contextual Risk Scoring Module (CRSM) enables priority-based AV decision-making for safety-critical unusual objects. DOI: DOI : https://doi.org/10.52783/pst.3468

Links and sources

• PDF from original source

Comments