摘要详情

ID / 提交时间

61 / 2023-08-30 20:23:17

标题

Production line action behaviour recognition based on dynamic attention mechanism in sequence space

关键字

deep learning, video behavior recognition model, spatial masking

主题及专题

Special Sessions > SS11 - Human Activity and Scene Understanding

状态

终稿

作者

Hongsheng Li / Xidian University

Han Bing / Xidian university

Liang Zhang / Xidian University

摘要

Product surface integrity inspection is a key step to ensure product quality, and is usually done manually by quality inspectors. In order to reduce the mistakes caused by fatigue caused by repeating a single action for a long time, this paper intends to use an auxiliary surface integrity detection system based on deep learning to remind and check whether it is complete. However, data-driven deep learning models usually require a large amount of data, and there are few action datasets in industrial scenes. Most of the existing video datasets are mainly concentrated in the field of daily life and cannot be directly applied to industrial scenes. To address this issue, we propose Surface Integrity Check Dataset (SIC), a new large-scale dataset for human behavior understanding in industrial scenarios. The SIC dataset provides a series of action videos of quality inspectors performing product surface integrity inspections under industrial scene conditions, specifically showing whether the quality inspectors perform a complete spatial 360° inspection of the product. The difference between the SIC dataset and the general action understanding dataset lies in its action background, the correlation of action categories and its multi-view feature. The SIC dataset contains 9183 31-category object appearance inspection samples and 2 action categories, and for each video clip, we provide videos from three perspectives. At the same time, an action behavior intelligent detection model Mask-X3D based on deep learning technology is proposed. This model uses a large convolution kernel for efficient feature extraction, adaptively removes irrelevant background information in the scene, and conducts experiments on the self-built SIC dataset. The experimental results show that the recognition accuracy of the model proposed in this paper is 93.82% on the self-built dataset, which is better than 91.29%, 91.73% and 93.57% of the existing I3D, SlowFast and X3D networks, meeting the field accuracy and Lightweight requirements can effectively monitor the standardization of inspection actions, thereby improving product quality and production efficiency, and reducing risks such as unqualified packaging and product returns.