AI-Driven Object Detection Using YOLOv7 and Enhanced Small-Object Recognition with Vision Transformers for Real-Time Surveillance Solutions

The aim of this study is to create an object identification framework that enhances current models by utilizing YOLOv4, YOLOv6, YOLOv7, and Vision Transformers. The main objectives are to improve computational efficiency, small object recognition, and real-time detection capabilities. Materials and Methods: The experimental setup consists of two groups: Group 1 (Control) utilizes YOLOv4 and YOLOv6 on the PASCAL VOC dataset, comprising 16,551 training and validation images. However, it exhibits limitations in accuracy and scalability and Group 2 (Intervention), refers to 200,000 Images and 80 object categories, which introduces a novel technique that enhances small and occluded object detection with a 95% confidence interval through the integration of Vision Transformers and hybrid architectures utilizing YOLOv7. Results: The proposed YOLOv7 + Vision Transformers framework achieves higher accuracy than YOLOv7. Its precision ranges from 94.8% to 95.0% with accuracy 94.5%, outperforming YOLOv7's 91.2% to 93.1 %. The optimal inference time is 20.8 ms, significantly lower than YOLOv7's 21.1 ms to 23.9 ms, with a significance of 0.05.Conclusion: In this work, its is observed that the YOLOv7 and Vision Transformer has significantly provide a better accuracy and establishing a robust solution for enhanced detection capabilities.