Real-time speed bump detection using YOLOv8 and embedded vision for adaptive braking assistance in electric vehicles

Abstract

There are many unmarked speed bumps on our roads, which cause discomfort and damage to EVs and cars, including body damage, shock to occupants, and hard braking of the electric motor. This paper proposes a real-time embedded system for speed bump detection and a corresponding algorithm to activate the intelligent braking system of an EV, implementable in even the lowest grade of Advanced Driver Assistance Systems (ADAS). The base model used is YOLOv8-small, trained using transfer learning weights on a custom dataset of approximately 2000 images scraped from Mendeley and Kaggle, covering a wide range of bump shapes and sizes. The images are annotated in Roboflow by drawing polygons for the class "speed bump." The model is trained on Google Colab GPU for approximately 150 epochs and implemented on a Raspberry Pi 4 with 8 GB of RAM and a Pi Camera V2. The vehicle braking is controlled via a PWM signal through the L298N H-bridge driver, with GPIO-based audio-visual alerting also implemented. The YOLOv8 model achieved a precision of 0.9777, recall of 0.9977, mAP@50 of 0.9958, mAP@50–95 of 0.8577, and an F1 score of 0.99. On the Raspberry Pi 4, the system achieves a raw inference rate of 0.84 FPS (~1522 ms) and effective decision pipeline throughput of 8–10 FPS via frame-skipping. Model performance on a Mac M3 GPU achieved 43.03 FPS. The proposed system is highly scalable and cost-effective, deployable as a low-cost ADAS module in EVs to improve passenger safety and extend vehicle life across varied road surfaces.