Prevention of Motorcycle–Car Door Collisions by Using a Deep-Learning-Based Automatic Braking Assistance System

Collisions between motorcycles and car doors that are being opened are common, preventable accidents that can result in fatalities. A critical limitation of safety advancements in both cars and motorcycles is high cost associated with the use of radar sensors. In this study, a deep learning model was integrated into an inexpensive and camera-utilizing automatic braking assistance system for motorcycles to enhance braking performance and alert motorcyclists to avoid collisions. This research involved two stages: (1) the training of a deep learning model for detecting car door states and (2) the design of safety mechanisms for selecting appropriate braking intensity and front braking ratio values on the basis of the model's output, time-to-collision, the rider's braking action, and the initial braking speed, in order to achieve optimal braking performance. Specifically, the YOLOv12s object detection model showed high performance in predicting the states of car doors, exhibiting precision, recall, and mean average precision values of 90.5%, 80.6%, and 87.8%, respectively. The braking intensity of the system was set to 0%, 25%, 50%, or 100% in scenarios involving opening states of the car door (closed, small, medium, or large opening), time-to-collision values, and the rider's braking action. The optimal front braking ratio function was determined based on the initial braking speed to achieve the optimal braking performance. At an initial braking speed of 60 km/h, the braking stroke under a front braking ratio of 45% was 35.61% and 13.37% shorter than those under front braking ratios of 20% and 60%, respectively. The proposed braking assistance system can feasibly be deployed in the real world because it can respond within a safe time window under the conditions studied, which is approximately 0.5 s. However, further refinement is required, including improvement of the robustness of the object detection model through the collection of a larger and more diverse dataset, experimental measurement of front braking ratios to determine the optimal braking performance in real scenarios, and design of a physical actuator to control braking intensity and the front braking ratio in real time.