Real Time Detection and Warning System for Monitoring Pedestrian Traffic on Crosswalks

Pedestrian fatalities on the road have been steadily increasing over the past decade with most accidents occurring at night. Many current methods for pedestrian detection are costly and require expensive computer hardware to monitor intersections on roadways. In this thesis we propose a embedded pedestrian detection system was developed using a microcomputer and a lightweight convolutional neural that was capable of performing computer vision inferences on small memory devices that could be easily installed at intersections. The neural network was trained on a custom made dataset of images that were captured with a camera that could switch between a standard camera and a infrared cut camera that is sensitive to infrared light at night making night time detection possible on a embedded detection system. The detection system was designed to monitor a two-lane roadway (24ft) in order to continuously monitor crosswalks for pedestrians. Once a pedestrian was detected by the detection module a signal would be transmitted to the warning module which consisted controlled a flashing beacon that could be used to warn on coming drivers. The system was able perform both during the day and at night in pitch darkness. The system was able to achieve a overall accuracy of 95.37%. The system was tested and validated at the University of Texas at San Antonio campus.