Laboratory Dangerous Operation Behavior Detection System Based on Deep Learning Algorithm

Introduction

Human behavior is an important part of human life. There are many kinds of hazardous chemicals in university laboratories, which are scattered in storage locations and densely used (Dewi, 2021) and they are flammable, explosive, toxic, infectious and corrosive, and some high-risk chemicals even have highly toxic characteristics. If we can quickly identify dangerous behaviors by monitoring the video taken by the camera in the laboratory, we can find potential accident risks in time and deal with them immediately, which can effectively reduce or eliminate laboratory safety problems and ensure personal and property safety. Behavior detection and recognition has attracted more and more attention from relevant researchers because of its great application potential in monitoring system, video analysis and other fields (Ophoff et al., 2020; Li et al., 2021), and how to solve various problems in video behavior detection and recognition by using DL (Deep learning) technology is the hottest topic (Chen et al., 2021; Xu et al., 2020).

At present, the target detection methods based on DL algorithm can be divided into two categories according to whether the regional candidate network is adopted or not. The first stage is the two-stage target detection algorithm, the first stage is to generate candidate regions through the regional candidate network, and the second stage is to classify and regress the candidate regions (Xing et al., 2020). Literature (Zheng et al., 2021) put forward the histogram feature of gradient direction, combined with simple linear support vector machine as classifier, and achieved the best effect in the human detection algorithm at that time. Literature (Dao et al., 2022) proposes to use multi-camera pictures to detect the position, gesture and finger bending of hands, and stable detection can be realized through bone pictures, instead of data gloves, a contact device. The slow fusion model proposed in reference (Zimoch & Markowska-Kaczmar, 2021) uses 3D convolution and average pooling in its first three convolution layers, and achieved the best behavior detection effect at that time. Literature (Kohiyama & Yamashita, 2020) proposes a feature pyramid network for small target detection on the basis of Faster RCNN. Before using the feature pyramid structure, most DL-based detectors only detect at the top of the network.

Over the years, China has carried out a lot of research work in sensor network and monitoring, but the research combined with laboratory site safety monitoring is still in its infancy and exploration stage. Therefore, in order to manage hazardous chemicals in university laboratories scientifically, normatively and efficiently, it is particularly important to build an information management platform covering the whole process of purchase, use, storage and abandonment. In this paper, a multi-sensor integrated, flexible combination of functions, compact appearance, high precision and low cost laboratory dangerous operation behavior detection system is designed to improve the efficiency of laboratory safety management.Traditional methods for detecting hazardous work behaviors often rely on manually designed feature extraction methods, which often struggle to achieve ideal results in complex and dynamic environments. However, the emergence of deep learning algorithms provides new solutions to this problem. By constructing deep neural networks, we can automatically learn and extract high-level features from images, thereby achieving more accurate behavior detection.

In this article, we adopt a network structure of deep learning neural networks. This network structure can effectively capture spatial and temporal information in images. Firstly, deep learning is used to extract spatial features from raw images to identify the actions and postures of personnel in the laboratory. Then, deep learning is used to process temporal information to identify patterns of continuous actions and predict possible future behaviors. This network structure not only improves the accuracy of detection, but also enhances the real-time performance of the system.