Enhancing the Accuracy and Speed of Object Detection and Distance Estimation to Improve the Safety of Autonomous Cars Movement

Abstract. Modern transportation systems face significant challenges in ensuring road safety, with approximately 35.1 million fatalities annually due to accidents, 93.5% of which are attributed to human errors. Autonomous vehicles have the potential to mitigate these incidents. They are classified into six levels of automation, from none to full automation, with obstacle detection and distance estimation being fundamental across all levels. This paper focuses on level 1 automation (driver assistance) in Iran, where most of the vehicles currently operate at Levels 0 and 1. It utilizes colour images captured by an SM-A52 mobile camera (2084 × 4624 pixels) in Districts 6 and 11 of Tehran under varying environmental and traffic conditions.
To enhance accuracy and speed in obstacle detection, four YOLO algorithm versions were implemented, with YOLOv8-s selected for its superior performance based on mean average precision, recall, and processing speed. For distance estimation, stereo imaging with two mobile cameras placed one meter apart was employed. Calibration parameters were obtained, and a 3D model was generated using Structure from Motion to calculate distances. The results were evaluated using Mean Absolute Error and Root Mean Squared Error, achieving a 20% increase in accuracy for obstacle detection compared to previous studies. Despite using more limited equipment, this research achieved comparable accuracy with respect to earlier works.