DynaSeedFusion: A real-time object-level dynamic obstacle detection system for agricultural robots

Reliable dynamic obstacle detection is essential for agricultural robots in unstructured farm environments. Light detection and ranging (LiDAR) returns are often sparse, and vegetation and fence boundaries, partial occlusion, and slow or intermittent motion further challenge lightweight methods based on instantaneous velocity thresholds. This paper presents DynaSeedFusion, a real-time system that combines motion-seed-guided candidate formation with object-level evidence aggregation and short-term decision confirmation. Motion seeds guide LiDAR candidate formation, red–green–blue-depth (RGB-D) observations provide visual and geometric support, and short-horizon temporal consistency describes the level and stability of recent motion. The three evidence terms are combined using confidence-proportional weighting. A pretrained YOLOv11n (You Only Look Once, version 11 nano) detector is used without farm-specific retraining. DynaSeedFusion is evaluated in two stages on three datasets, covering 27 sessions and 1,003 manually annotated evaluation frames collected using the same robot. The first stage uses the Farm-like Set and Open Farm Set; the second uses the In-Farm Dataset, which was collected at a cattle farm and contains 16 sessions and 488 frames. Human participants were used as controlled moving targets in these trials. On the In-Farm Dataset, DynaSeedFusion achieves an F1-score of 0.741 at the primary three-dimensional intersection-over-union (3D IoU) threshold of 0.3, compared with 0.649 for LiDAR–Visual Dynamic Obstacle Detection and Tracking (LV-DOT). At IoU = 0.5, scores are 0.335 and 0.230. On the same motion-seed-guided candidate set, the LV-DOT velocity-and-voting classifier gives an F1-score of 0.531, while the proposed decision method raises it to 0.741. On a 55 W onboard mini PC, the C++ implementation requires 33.4 ms per LiDAR frame on average and supports the 10 Hz sensor update rate.