Development of a collision warning system for personal mobility devices applications

Abstract

The Personal Mobility Devices (PMDs) are gaining their soaring popularity worldwide. In Singapore, they serve as a significant option for the short trips as well as the first- and -last mile journeys. However, though several measures such as speed limit had been taken for safety consideration, the number of accidents involving PMDs was still nonnegligible. To deal with this threat to both riders and pedestrians, this research develops a vision-based collision warning system specifically for the PMDs applications. During the development process, several candidate hardware configurations are comprehensively compared and the system consisting of OAK-D, Raspberry Pi, Arduino Nano and actuators is finally proved as the most cost-effective. MobileNet-SSD model is employed for object detection and the test results show the errors caused by negative influencing factors are within an acceptable range. Moreover, the officially recommended disparity confidence for depth estimation is further optimized based on the relative error and standard deviation of the measurement results. With specifications of the common PMD products and field experiment data, a warning mechanism for different collision risks is calculated and established. An online survey to collect the subjective safety assessments for riding state is also conducted and a transition zone with a haptic warning for riders only is added to the previous model for the consideration of better user experience. Compared with previous multi-sensor-based designs, this system achieves a better balance between cost and overall performance with a more concise and efficient structure since all the environment perception tasks are performed relying solely on one powerful camera. Besides, combining the riders’ feeling with the classic safety distance approach is also innovative. For the future work, a more targeted neural network trained with the real traffic data of Singapore, further detailed investigations on quantifying the subjective feelings and a better collision prediction model with velocity involved are preferred.