AI model detects overtaking cars from bike cameras with 97.8% recall

A team led by Gandhimathi Padmanaban at the University of Michigan has built a computer vision pipeline that automatically detects overtaking vehicles from a single rear-facing bicycle camera. The system combines RT-DETR object detection with ByteTrack multi-object tracking, then applies a three-stage geometric validation module that checks bearing angle trends, apparent size growth, and spatial confirmation—all derived from perspective projection principles. This eliminates the need for multi-sensor setups or explicit camera calibration, and removes the manual annotation bottleneck that has limited naturalistic cycling safety research.

Validated on 315 real-world overtaking events from urban roads in Ann Arbor, Michigan, the pipeline achieved 97.8% recall with zero false positives. It detects overtakes a mean of 2.44 seconds before the vehicle passes, with 84.1% of events coming more than 1.5 seconds before passage—enough time for an active cyclist warning. The system also estimates lateral passing distance using bounding box geometry, achieving mean absolute errors of 13-14 cm, enough to distinguish close passes (under 5 feet) from standard ones. Notably, 33.3% of passes in the test set were below the 5-foot threshold, consistent with prior studies. This work provides a scalable, calibration-free approach to automated cycling safety analysis from consumer-grade footage.