Zero-Shot Pedestrian Detection in Autonomous Driving Systems under Day vs Night Conditions

NLDL 2026 Conference Submission62 Authors

Keywords: micromobility, zero-shot pedestrian detection; low-light detection; BDD100K; RT-DETR; RT-DETRv2; transformer-based object detection

TL;DR: Zero-shot pedestrian detection on BDD100K across day/dusk/night with RT-DETR(v2): night recall drops sharply; v2 > v1. A 0.5 score threshold works best, motivating night-focused adaptation for ULEVs.

Abstract: Autonomous ultralight vehicles operate in varied lighting conditions, where robust pedestrian detection is critical. This paper examines zero-shot pedestrian detection performance across day, twilight, and night scenarios using modern Transformer-based models. We leverage the large-scale BDD100K driving dataset to compare Real-Time DEtection TRansformer (RT-DETR) against its improved successor RT-DETRv2 on identifying pedestrians without any fine-tuning. Our experiments fix IoU at at 0.5 and analyze recall as detection confidence varies. Results show a significant drop in recall from day to night, indicating that low-light conditions degrade detection. RT-DETRv2 consistently outperforms RT-DETR, recovering a portion of missed detections under all lighting conditions. We discuss the implications for deploying these models in ultralight electric utility vehicles (ULEVs) where human operators and vehicles share tasks, highlighting the need for adaptive learning and operator feedback to maintain safety after dark. Future work will integrate interactive learning to improve night-time perception.

Serve As Reviewer: ~Faisal_Shafait1, ~Marcus_Liwicki1, ~Foteini_Simistira_Liwicki1, ~Nudrat_Habib1, ~Tosin_Adewumi1, ~Adrian_Ulges1, ~Ulrich_Schwanecke2

Submission Number: 62