Optimization of YOLOv4-Tiny Algorithm for Vehicle Detection and Vehicle Count Detection Embedded System

Abstract

Currently, the implementation of object detection systems in the traffic sector is minimal. CCTV cameras on highways and toll roads are primarily used to monitor traffic conditions and document violations. However, the data recorded by these cameras can be further utilized to enhance traffic management systems. The author proposes a vehicle detection and counting system using YOLOv4-Tiny. The research aims to improve vehicle detection and counting accuracy by employing a median filter and grayscale processing, which simplify object detection. The proposed YOLOv4-Tiny algorithm has shown impressive results on various datasets, including MAVD, GRAM-RTM, and author dataset. The system achieved a detection accuracy of 98.95% on the MAVD dataset, 99.5% on the GRAM-RTM dataset (comparable to YOLOv4), and 99.1% on the author dataset. Furthermore, the system operates at 25 frames per second (FPS), a notably high rate compared to other methods. While the system demonstrates excellent accuracy in counting cars, it encounters some accuracy loss with other vehicle classifications. The author concludes that the system is highly suitable for real-world applications but notes that inaccurate labeling can lead to vehicle counting errors.