A Low-Budget, End-To-End Warning System for Bicycles using Monocular Vision and Vibrating Handlebars

Awareness of surrounding traffic is crucial for safety. Nowadays warning and information technology in cars is often used to supplement people’s awareness of their surroundings. This thesis presents such a warning system, but designed for bicycles instead. An end-to-end, low budget system was developed, that warns cyclists of upcoming traffic. The only sensor input to the system is a monocular camera, and the processing is done on a cheap 100$ computer. For processing the choice was made to use the Yolov4 neural network, in combination with custom algorithms for tracking and finding the direction of travel. Additionally, a survey was performed to explore people’s acceptance of such a system, and find the necessary performance such a system must have. The survey indicated such a system must have warning times of at least 3-4 seconds, which in practice was only reached less in less than 50% of cases. Additionally parked cars often created false warnings. The main obstacle in having a faster warning time or less false positives was the performance of the tracking algorithm.