Incremental Learning-Based YOLOv5 Detector for Efficient Labor Protection Products Detection

초록

In recent years, the accident rates remain high in the construction industry, and the construction safety is a matter of great concern worldwide. An intelligent and efficient incremental detector that can accurately detect whether the workers properly wear labor protection products, such as Safety helmet, safety belt, reflective clothing, is crucial for construction safety. Currently, the most advanced incremental detectors mostly adopt external fixed region proposal methods based on knowledge distillation, which require a lot of time and cost. Due to the lack of annotations for old categories and region proposal information in single-stage detectors, the detectors usually identify old category targets as background, leading to catastrophic forgetting. To address the above problems, this paper proposes a data label merging algorithm, which uses the detection results of the old model and existing data labels to generate merged labels for training, which compensates for the missing annotations of old category targets in the new dataset and mitigates catastrophic forgetting. At the same time, a recent mean-based exemplar selection algorithm is proposed to select representative samples stored in the old dataset, and this representative sample set will be used for model training to help the model resist forgetting. This method has been implemented in a single-stage detection framework, and its effectiveness has been verified on the LPAD (Labor Protection Appliance-Dataset) dataset. Our proposed method improves the average precision score (mAP) of old categories and total categories detections by 6.9% and 6.0% respectively compared with the current state-of-the-art method. The merged labels and representative samples obtained from the proposed method effectively alleviate the forgetting problem. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024.

키워드