Paper Key : IRJ************647
Author: Bommana Bhavana
Date Published: 02 May 2024
Abstract
Effective highway management systems require intelligent vehicle recognition and counting; however, there are still issues since different vehicle sizes affect detection accuracy. This study describes a vision-based system for vehicle recognition and counting that uses input video streams to create YOLOv8 with lessons from YOLOv7.A unique segmentation approach is presented to extract the highway road surface from video frames, addressing the problem of different vehicle sizes. By concentrating on pertinent regions, this segmentation improves detection accuracy. The separated areas are then processed with YOLOv8, which incorporates improvements from YOLOv7, to efficiently identify the kind and location of vehicles.Post-processing techniques and a maximum suppression algorithm are used to modify detection findings in real-time, improving precision, especially in situations where there are overlapping or tiny cars. Additionally, vehicle trajectories are obtained from identified vehicles, allowing for the identification of various vehicle kinds and the estimation of driving directions. To confirm the efficacy of the suggested methodologies, a variety of highway surveillance movies are used for experimental assessments. The results show noteworthy progress in determining driving directions and counting cars, as well as considerable increases in the accuracy of detecting tiny vehicle objects.Keywords: vehicle counting,yolo, post-processing step, non-maximum suppression, simple deep convolution.Counting and detecting vehicles has become a critical problem in several fields, including traffic management, surveillance, and autonomous driving. Typical techniques frequently have issues with speed, accuracy, and vehicle type recognition. Advanced algorithms like "You Only Look Once" (YOLO) have become more popular as a solution to these problems. YOLO achieved great accuracy and real-time performance, revolutionizing object identification. In contrast to conventional techniques, which require many processing steps, YOLO effectively makes predictions by processing the complete image through a neural network in a single pass. The basic idea behind YOLO is to anticipate bounding boxes and class probabilities for each grid cell in the input picture by first splitting it into a grid. Vehicle localization and categorization may be done precisely with this grid-based method.In this project, we aim to implement a neural network-based vehicle detection and counting system using YOLOv8, which builds upon the strengths of previous versions while integrating novel features and optimizations. By leveraging the efficiency and accuracy of YOLOv8, we endeavor to address the challenges associated with vehicle detection and provide a robust solution for real-world applications. Subsequent iterations of YOLO, such as YOLOv2, YOLOv3, YOLOv4, and YOLOv5, have been introduced in order to improve vehicle detection capabilities and overcome limitations in earlier iterations.Accurate vehicle identification and counting is essential to many applications in contemporary transportation systems, from intelligent transportation systems to traffic monitoring. The accuracy and effectiveness of traditional vehicle identification techniques are frequently compromised, especially in difficult situations like congested traffic areas or low light levels. Neural network-based techniques, on the other hand, have completely changed this sector by providing improved capabilities for reliable and fast vehicle recognition and counting.
DOI Requested