A semantic segmentation-based collision recognition method for motorcycle slalom through poles in Motorcycle Driving License Test

Abstract

Motorcycles are one of the most important means of transportation. For traffic safety, most countries require drivers to pass the Motorcycle Driving License Test (MDLT). The traditional MDLT relies on manual assessment, leading to expensive labor costs, inconsistent test standards, and unsupervised processes. Therefore, the intelligent MDLT for automatic assessment becomes an urgent need. This paper proposes a collision recognition method using semantic segmentation for motorcycle slalom through poles in the intelligent MDLT. The method identifies the pole and calculates the pole angle change according to the real-time video provided by the on-site camera. A collision between the motorcycle and the pole is recognized when the pole angle change is larger than the preset value. Specifically, we propose a Fast Flow Alignment Module (FFAM) to improve the efficiency of the semantic segmentation network. Then we build a lightweight semantic segmentation network using FFAM. Finally, we design a post-processing method to calculate the angle change value of the poles. Extensive experiments on a newly collected dataset of slalom poles demonstrate that our proposed network achieves a state-of-the-art trade-off between segmentation accuracy and inference speed: 77.6% mIoU and 167 FPS. Moreover, our post-processing method can accurately identify the pole angle with an error within ±0.1°. Our method is an essential component of the intelligent MDLT, which is fast, accurate, and has been successfully applied in many cities. Our video demo is shown at https://www.youtube.com/watch?v=gjE3Imne240.