Advanced Domain Adaptation Technique for Object Detection Leveraging Semi-Automated Dataset Construction and Enhanced YOLOv8
Ahmed Gomaa
Abstract
Detecting moving objects is a pivotal research do-main within Intelligent Transportation Systems (ITS) and various computer vision applications. This paper presents an innovative and resilient semi-automatic method for unsupervised object detection, which combines a customized adaptation of the CNN-based You Only Look Once V8 (YOLOv8) approach with background subtraction. First, background subtraction using low-rank decomposition is employed to detect moving objects. Next, a clustering method is utilized to refine the outcomes of the background subtraction process. These refined results are then employed to fine-tune the modified YOLOv8, which is subsequently used for object detection and classification. The primary contribution of this work lies in its novel detection framework that eliminates the need for manual labeling by creating an automatic labeling system. Extensive experiments carried out on practical object monitoring benchmarks reveal that the proposed framework notably enhances mean Average Precision (mAP) when compared with recent approaches on both the UA-DETRAC and CDnet 2014 datasets.