A stereo-vision system for real-time person detection in ADAS applications using a fine-tuned version of YOLOv5

Abstract

Pedestrian detection holds significant importance in advanced driver assistance systems (ADAS) applications, and presents a challenging task in this field. While the advent of deep learning has facilitated the introduction of various pedestrian detectors characterized by accuracy and low inference speed, there persists a need for further improvements. Notably, ADAS requires accurate detection of pedestrians in various environmental conditions that can adversely impact the model’s performance, such as poor lighting, and bad weather. Furthermore, an imperative requirement involves the incorporation of distance estimation in conjunction with pedestrian detection, with an extension of detection capabilities to encompass cyclists and riders, who are equally crucial for ensuring road safety. Therefore, this paper introduces a stereovision system designed for the detection of pedestrians, cyclists, and riders. The initial phase, involves improving the performance of you only look once version 5 (YOLOv5s) through a fine-tuning process with a custom dataset integrating augmentation techniques to common objects in context (COCO) dataset. The detector is trained using Google Colab, and tested in real-time with a Raspberry Pi 4 model B, 8 G RAM. A comparative analysis is conducted between the YOLOv5s and the fine-tuned model to prove the accuracy of our approach. The results showcase a high performance of the detector reaching an accuracy exceeding 79%.