In recent years, deep learning-based human skeleton action recognition technology has seen significant development and application in areas such as human-computer interaction, security monitoring, and rehabilitation medicine. To address the issue of insufficient exploitation of spatiotemporal features in skeleton action recognition, this paper designs a Spatiotemporal Graph Convolutional Network with feature enhancement. This network enhances features from the perspectives of nodes, space, and time to improve the accuracy of skeleton action recognition. Firstly, by connecting graphs of adjacent frames to construct local spatiotemporal graphs, the network achieves a more comprehensive capture of spatial and temporal dependencies. Next, in the spatial graph convolution operations, self-attention mechanisms and channel attention modules are used to learn separate topological structures for different time frames, allowing the model to capture complex changes in human body structure across various time frames. Finally, in the temporal convolution stage, the use of multi-scale convolutional kernels enhances the perception of features across different temporal domains. This human skeleton action recognition model demonstrates efficient spatiotemporal feature extraction capabilities and excellent performance.
At present, indoor positioning technology generally has some problems such as the need for pre-deployment of equipment, poor anti-interference ability and high cost. To solve these problems, this paper designs a lightweight network model based on deep learning technology, and uses pure visual method to deduce indoor positioning coordinates according to indoor scenes. Firstly, since there is no open data set for visual positioning of Indoor scenes at present, an Indoor Dataset 600(ID600) dataset was built by ourselves. Then, different lightweight network models were trained on SF-XL(Small) outdoor geographic image positioning data set to investigate the positioning accuracy and reasoning speed of different models. Finally, the structure adjustment and training strategy of the model are optimized on ID600. After optimization, the number of parameters in the Local-SqueezeNet model is only 1.14M, and the Top-1 accuracy reaches 91.4% when the Distance threshold is set to 1200mm.
In order to ensure the safe and efficient operation of subway stations, it is necessary to obtain air quality data and passenger flow information in subway stations in real time. In this paper, an intelligent monitoring system for subway environmental safety has been built, which consists of three modules: subway air quality detection module, subway passenger flow detection module and data display storage analysis module. Particulate matter (PM), CO2 and formaldehyde which are hazardous to the health of passengers and staves has been detected by subway air quality detection module. The subway passenger flow detection module can obtain on-site video, utilize object detection algorithms to acquire real-time passenger flow information in subway station. All data is displayed, stored and analyzed through the local wireless LAN upload server. Through the data interaction between each edge device and the server, air quality data and passenger flow information in the subway environment can be monitored in real time, so as to realize the digitalization, precision and intelligence of subway operation management.