Computer Vision Driven Automated Bicycle Braking
| December 2025 | University of California San Diego |
Summary
OptiBrake is a low-cost bicycle collision-avoidance system: a front camera runs Lucas-Kanade optical flow to detect approaching obstacles; a closed-loop controller runs the brake motor until wheel speed hits zero, then reverses to unspool the cable.
Motivation
E-bikes and mixed traffic increase collision risk; automotive AEB usually needs LIDAR/ultrasonic and is expensive. We explored monocular vision (optical flow) plus wheel odometry for distance—detecting any obstacle without classifying it—and showed that single-camera, Pi-based braking is feasible in controlled conditions, while documenting limits (e.g., false positives outdoors).
Contributions
I worked on wheel odometry, closed-loop brake control, and hardware integration. I also helped design and analyze the indoor trials, and debugging the optical flow implementation, and it’s integration with our control loop.
Stack
- Language: Python
- CV: OpenCV, Lucas-Kanade optical flow (no neural network)
- Hardware: Raspberry Pi, gpiozero; Sony IMX500 camera; Keyestudio sensor + spoke cards; 12V gear motor (150 RPM), L298N, V-brakes, dual batteries
- Method: Focus of Expansion and time-to-collision with wheel odometry for distance
Challenge & approach
Problem: Single-frame optical-flow thresholds gave okay latency but lots of brief false positives (trees, buildings, bike jitter), especially at low speed and outdoors.
Approach: Debounced wheel ticks (register only if time since last “real” tick exceeds a threshold), 3 ticks/rotation, and a weighted average of the last 9 RPM samples (current ×5) for a stable speed signal. We ran controlled indoor trials at matched speed to validate braking and distance-to-collision, and documented that outdoor robustness would need more compute or extra sensors (e.g., LIDAR).
Highlights
- Closed-loop braking — Wheel speed as process variable; software controller; motor actuates until RPM = 0, then reverses to unspool.
- Interrupt-based wheel speed — Debounced ticks and weighted average of last 9 RPM; no ABS (DC motor + cable can’t do fine PWM).
- Optical flow detection — Lucas-Kanade for any-obstacle proximity; FoE and TTC with odometry for distance.
- Indoor validation — Showed automated brake actuation from single-camera CV on a Pi in controlled trials; documented false-positive and environment limits.
