Safety for all road users is a key concern as Connected and Autonomous Vehicle technologies develop and reach the testing phase. One key concern is the change to two-way communication that often occurs in traditional pedestrian-vehicle interaction. This project focused on three key aspects to this issue. First, a prototype autonomous shuttle system was expanded to include additional communication features in a lightbar as well as more advanced pedestrian detection systems. Secondly, multiple detection methods were trained and tested using traditional datasets as well as a new dataset including occluded pedestrians. Finally, a survey was conducted to determine how well pedestrians understand specific static or dynamic lightbar patterns as an additional communication tool for CAVs.
This project tested multiple pedestrian detection methods and developed improved methods with increased accuracy and reduced latency. The EcoPRT vehicle was able to incorporate the improved detection method, however the training image set included multiple camera perspectives and the method could likely be applied to infrastructure-based detection systems. Additionally, the project developed a body part-based method which detects head, arms and legs of pedestrians in order to improve the overall detection of pedestrians when they are partially occluded. The project also developed a database of occluded pedestrian images which can be used for training or testing other new methods addressing this issue. Finally, the project examined multiple methods for signaling the CAV intent to pedestrians using fixed or moving lightbars. Respondents struggled to correctly identify the message communicated by the lightbar in cases where multiple movements are expected (such as locations with potential turning movements) but identification improved in more constrained environments.
