Improvement of YOLOv8 Detection Algorithm for Worker-Related Objectives in Construction Scenarios

Abstract

In high security risk construction scenarios, the wearing of personnel safety equipment can effectively reduce safety risks. The performance of general YOLO series algorithms is still not sufficient when they are used in construction scenes to detect objects, such as pedestrians especially workers, helmets and reflective vests. This paper further optimizes the detection algorithm on the basis of YOLOv8 algorithm. The main contributions are as follows: the detection heads are modified to enhance the ability to extract small objects, a location loss function is replaced to optimize the model training process, and the attention mechanism module is added to better extract crucial features. This paper mainly conducts experiments and evaluations based on dataset SODA-C3 and dataset GLD-HR made for construction scenarios. The mean average precision of the optimal model trained by the improved algorithm reaches 83.1% and 88.5% respectively, which is 6.3% and 2.7% higher than that of the benchmark model, and the inference time is less increased.