부산대학교 도서관

This paper presents a vision-based approach for detecting speed bumps, which is crucial for enabling safe and efficient speed control in autonomous vehicles. Given the diverse range of speed bump sizes and characteristics encountered in Indian scenarios, a robust detection algorithm is required. To this end, we evaluate two state-of-the-art deep learning based object detection models, Faster R-CNN and YOLOv5, and compare their performance. Our study specifically focuses on detecting both marked and unmarked speed bumps in real world environments. However, we also address the challenge of misclassifying pedestrian crosswalks, which can be mistaken for speed bumps due to their similar features. To enhance the accuracy of detecting marked speed bumps, we employ the Negative Sample Training (NST) method. The results show that training with NST improved the detection performance of both Faster R-CNN and YOLOv5 models, achieving an average precision increase of $ 5.58\%$ and $ 2.3\%$, respectively, for marked speed bump detection. Furthermore, we conduct real-time testing of the proposed model on the NVIDIA Jetson platform, which yields an inference speed of $18.5\mathrm{~ms}$ per frame.