Lightweight Vit-Based Pedestrian Multi-Object Detection in Building Construction Scenarios

In the construction environment, significant safety hazards and pressing safety detection requirements are evident. Multi-target detection technology plays a pivotal role in ensuring real-time monitoring of safety and regulatory compliance at construction sites. Nevertheless, the intricate nature of the construction environment, characterized by cluttered backgrounds, diverse personnel, and challenging equipment deployment conditions, poses formidable challenges to the efficacy of multi-target detection algorithms. To address these obstacles, we propose a detection methodology that integrates lightweight Vision Transformer (ViT) and lightweight Feature Pyramid Network (FPN). The Lite Vision Transformer (LVT) is a streamlined backbone network that effectively reduces network parameters while maintaining high detection accuracy by optimizing convolutional self-attention and recursive atrous self-attention mechanisms. Additionally, the Feature Pyramid Network (FPN) utilizes spatial multi-scale feature maps to enhance target detection precision. This approach excels in multi-scale target detection and identifying obscured objects, making it suitable for deployment in resource-constrained construction sites. Experimental assessments conducted on the COCO dataset reveal that our method reduces parameter count by approximately 60% compared to the YoloV3-Tiny model while improving accuracy by roughly 5% over traditional lightweight CNN networks.