Road Peculiarities Detection using Deep Learning for Vehicle Vision System

Abstract

Recent development of Advance Driver Assistance System (ADAS) has seen various advancement in object detection for vehicle vision system, particularly on the detection of other vehicles, pedestrians, road lane and signage. While these detections can provide assistant to avoid road accidents, they still lack to include road condition factors that also contributed to road accidents in Malaysia. This paper proposes a detection of the road peculiarities such as pothole and road bumps to act as additional safety feature in ADAS. With the breakthrough of deep learning in solving image recognition problems, this work takes advantage of Single Shot Detector (SSD)-MobileNetV2 as the detection algorithm, implemented on the real-time. The training images for potholes and road bumps taken from the Malaysia roads are fed into the detection model, and then the pre-trained weights are fine-tuned over the training process. The results show that the detection algorithm can predicts the potholes and road bumps, while exhibit the detection accuracy and confidence limitation due to the variety of shape and pattern of potholes and road bumps. Testing the detection algorithm with NVIDIA Jetson Nano yielded about 20 frames per second (fps), suitable for real-time applications.