Vehicle Detection for Cyclist Safety

The system is designed to help prevent collisions between cyclists and vehicles. It is optimized to work on a midsummer day in Tucson, Arizona. The deep neural network runs on energy-efficient hardware and can perform inference on images captured in real time. The network learned by training on a data set of pre-labeled example images. Using what it learned, the network is able to draw bounding boxes around objects it believes are cars and can report a confidence level for each prediction made. The physical system is composed of a weatherproof enclosure housing the inference hardware the neural network runs on, a battery, and a cooling fan. The camera is mounted on top of the enclosure under a clear dome, positioned to detect vehicles to the rear-left of the cyclist. The enclosure is placed on a bike rack on the back of a bicycle. Images are processed at a rate of at least eight frames per second and the entire system can operate for at least six hours.