Real-time direction-aware time-to-collision reasoning with pose-guided light detection and ranging grounding for pedestrian safety on autonomous ground vehicles

Near-field pedestrian–vehicle interactions pose a persistent challenge for low-speed autonomous ground vehicles (AGVs), as range-based heuristics cannot reliably distinguish approaching pedestrians from those moving laterally or retreating at similar distances. This paper presents a real-time framework for direction-aware time-to-collision (TTC) estimation and safety-aware pedestrian navigation on embedded platforms. Pose keypoints extracted by You Only Look Once version 8 Pose (YOLOv8-Pose) are associated with two-dimensional Light Detection and Ranging (LiDAR) measurements through field-of-view projection to obtain metrically grounded pedestrian states. This pose-guided LiDAR grounding strategy is referred to as the PLG framework. A constant-velocity (CV) Kalman filter then enables direction-aware TTC estimation, which is mapped to discrete safety states (SAFE/WARN/BRAKE) for closed-loop control. Experiments conducted on an NVIDIA Jetson AGX Xavier embedded platform demonstrate real-time throughput of 8–12 frames per second (FPS), a pose detection success rate exceeding 95%, and a mean camera–LiDAR bearing alignment error of approximately 2–3°. Evaluated on a dataset of 777 samples spanning person and non-person object interactions, the proposed framework achieves an overall safety-state classification accuracy of 99.74% and a BRAKE-class recall of 100%, compared with 36.68% accuracy and 80.65% BRAKE recall for a range-only baseline. These results demonstrate that direction-aware TTC reasoning substantially reduces false braking in lateral and retreating scenarios while preserving timely intervention for genuinely approaching pedestrians. Both the artificial intelligence (AI) model and its application to closed-loop safety control on physical hardware are demonstrated, yielding an interpretable and deployable safety framework for resource-constrained AGV platforms.