Reclining Public Chair Behavior Detection Based on Improved YOLOv5

Abstract

This study proposes an object detection algorithm based on the improved YOLOv5 network for the uncivilized behavior of reclining public chair, which often occurs in cities. The current object detection field is studied by a single object. For the behavior of a lying public chair, the object to be measured is composed of two parts: the chair and the human posture jointly. Furthermore, the features of the object will show a large variability under different shooting angles, so the model’s ability to extract features of the object is extremely important. This paper incorporates the Ghost module based on the YOLOv5 network to enable the model to learn more object features. The Ghost makes the neural network lighter by using linear convolution instead of nonlinear convolution, and its generated redundant features can help the model learn more object features and improve the model performance. In addition, this paper uses a new loss function EIoU to replace the original loss function CIoU. By comparison, EIoU solves the problem that CIoU fails in penalty terms under specific conditions. EIoU enables the model to converge faster and better. After experimental validation on the test set, it is shown that the improved YOLO network improves F1 by 3.5% and mAP by 4.2% compared to the original algorithm.