CPS: Medium: Smart Tracking Systems for Safe and Smooth Interactions Between Scooters and Road Vehicles

Project summary:

This Cyber-Physical Systems (CPS) grant will study smart tracking systems on scooters for ensuring safe and smooth interaction with other vehicles and pedestrians on the road. The smart system consists of inexpensive sensors, active sensing-based estimation algorithms, and deep learning-based robust image processing to enable trajectory tracking of all nearby vehicles on the road. If the danger of a scooter-vehicle collision is detected, an audiovisual alert is automatically provided to the car driver to make them aware of the presence of the scooter. The system also monitors the scooter rider's behavior, provides real-time feedback to improve rider compliance with traffic signals and sidewalk rules, and documents the information as a part of the rider's safety record. The key attractive features of the system are that it is inexpensive (< $500), is immediately useful on today's roads without requiring the vehicles on the road to be equipped with additional technology, and is potentially commercializable. This project contributes to society by improving safety of microtransportation systems, and broadens participation in computing via undergraduate research activities and promoting significant cross-disciplinary collaboration between faculty in engineering, computer science, and human factors.

The project will conduct research to develop two novel vehicle-tracking technologies. The two technologies, one based on use of a low-cost single-beam laser sensor and another based on a low-cost low-density Lidar sensor, can have applications in protecting vulnerable transportation users such as bicyclists, motorcyclists, scooter riders, users in developing countries, and also in other cyber-physical systems such as indoor robots. The computer vision system will handle rain, snow, and low lighting--which pose a major challenge by corrupting normal image data. New robust deep-learning-based recognition techniques will be developed that can effectively deal with corrupted image data sets. This will be achieved by novel nonlinear modeling of the low-complexity structures in both the clean data and in the image corruption using deep learning and allowing for mixed corruption types, varying severity, and possible corruptions in the training data itself. To ensure human-in-the-loop robustness, the project utilizes human subject studies to evaluate the effectiveness of a variety of audio and visual mechanisms for alerting the motorist and the scooter rider, including innovations such as providing visual cues of biological motion on the scooter to improve localization of the scooter by motorists.