Improving Safety in Autonomous Driving with the Use of AI for Object Detection Prediction

Abstract

In the last decade, object detection and machine learning-based algorithms have enhanced significantly, making self-driving vehicles a reality rather than a vision. Our research explores the advancements in machine learning-based object detection methods and their application to autonomous vehicle systems, specifically using the YOLO (You Only Look Once) algorithm. We begin with an overview of the YOLO algorithm, covering the YOLO architecture and comparing YOLO to alternatives like SSD and LiDAR. We then describe our case study in which we trained our own YOLO model using a custom dataset. Next, we analyze the results from the trained YOLO model to make conclusions about the YOLO algorithm. After conducting a literature review of dozens of old experiments to compare the YOLO alternatives to YOLO itself, we presented how YOLO is a better option for real-time driving scenarios in comparison to SSD (Single Shot Detector) and LiDAR (Light Detection and Ranging). In our case study of training a YOLOv8 model with our manually crafted dataset, the object detection accuracy for the model went from 20% to about 90% in only 50 epochs. We concluded that our research has highlighted YOLO’s power and high speed when it comes to driverless vehicle object detection, but we also acknowledged the room available for future improvements to make roads safer.