Pedestrian Behavior Maps for Safety Advisories: CHAMP Framework and Real-World Data Analysis

Greer, Ross; Desai, Samveed; Rakla, Lulua; Gopalkrishnan, Akshay; Alofi, Afnan; Trivedi, Mohan M. · 2023 · Unknown

DOI: 10.1109/iv55152.2023.10186648

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the limitations of vision-based pedestrian detection systems, which often fail in low-light conditions, during occlusion, or when pedestrians emerge suddenly. To mitigate these risks, the authors propose CHAMP (Crowd-sourced, History-Based Advisories of Mapped Pedestrians), a system that generates safety advisories by aggregating historical pedestrian data into high-definition maps. The motivation stems from rising pedestrian fatalities and the known ineffectiveness of Automatic Emergency Braking (AEB) systems in complex urban scenarios where visual cues are absent. The CHAMP framework operates on the principles of "Fleet" and "Repeat," leveraging repeated vehicle passes over geographic locations to infer pedestrian behavior patterns. The system consists of three stages: training, inference, and interaction. During training, the system aggregates pedestrian detections from video data with corresponding GPS coordinates, associating pedestrian counts with specific spatial intervals. For inference, it employs a ball tree nearest neighbor search to efficiently locate the nearest pedestrian hotspot relative to the ego vehicle. The interaction stage issues advisories if the distance to a hotspot falls within a calculated stopping distance threshold, which accounts for vehicle velocity, reaction time, and road friction. The authors evaluated the system using a dataset collected in La Jolla, California, comprising 10,000 video clips and 3 million frames annotated by human experts. Test drives were conducted in scenarios challenging for standard detectors, including blind turns, occluded pedestrians, and night-time conditions. Quantitative results demonstrate that CHAMP can successfully issue advisories in scenarios where direct detection fails. The system’s performance is heavily influenced by the sampling distance hyperparameter, which determines how frequently the system checks for nearby hotspots. A sampling distance of 2 meters provided the best balance of precision and recall, whereas larger distances (e.g., 5 meters) significantly reduced recall, leading to missed advisories. In specific test clips, such as those involving occluded pedestrians or night-time driving, the system achieved high recall rates (up to 0.75) at optimal sampling distances, confirming its ability to detect latent pedestrian zones. However, precision varied; densely populated areas resulted in lower precision due to false positives, while infrequently crossed intersections yielded zero recall due to insufficient training data. The significance of this work lies in its potential to enhance Advanced Driver Assistance Systems (ADAS) by providing preemptive warnings in safety-critical situations where AEB systems are less effective. By shifting from single-instance detection to historical aggregation, CHAMP offers a robust method for identifying pedestrian hotspots despite environmental challenges. The authors conclude that while the current system requires extensive data collection to minimize false positives, future improvements involving trajectory prediction models and time-of-day stratification could further refine advisory accuracy. This approach provides a viable pathway for developing informative HD map layers that support both autonomous driving and driver vigilance.

Key finding

The CHAMP system successfully generates pedestrian safety advisories in challenging conditions like occlusion and darkness by leveraging aggregated historical detection data, achieving optimal performance with a 2-meter sampling distance.

Methodology

on_road

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• vru facing ehmi