| Human activity recognition can be widely used in various pervasive computing applications,and in outdoor scenes,pedestrian intrusion detection with different task requirements can be performed through target detection and other methods.Indoor scenes can be used as a new type of human-computer interaction to perform intelligent medical monitoring,such as smart home or elderly fall detection,or can be applied to interactive somatosensory games by recognizing human behavior.With the rise and widespread application of deep learning,the real-time and accuracy of human activity recognition have also greatly improved.Most traditional human activity recognition systems use wearable sensors for detection and recognition.Although with the development of technology,the size of sensors has become smaller and data collection has become more effective,sensor-based detection systems still face the limitations of deployment.Especially with the continuous increase of sensing range and scale,the cost of deploying and maintaining large-scale sensing systems will also increase sharply.Therefore,passive human activity recognition without devices has gradually become the mainstream of research.Currently,passive human activity recognition methods without devices are mainly divided into two categories,namely,passive detection based on computer vision and passive detection based on radio frequency fingerprint signals.Passive detection based on computer vision uses a camera as a collection tool to detect human behavior through the collected image information,and is commonly used in outdoor pedestrian detection,face recognition,and other fields.Passive detection based on radio frequency fingerprint signals uses radio frequency signals as the sensing subject(commonly used radio frequency signals such as Bluetooth and Wi Fi),and is typically applied to human activity recognition in indoor scenes that require privacy assurance.This paper uses computer vision based methods and Wi Fi signal based methods to achieve the task goal of human activity recognition for outdoor and indoor scenes.In outdoor scenes,this paper conducts label processing and image preprocessing on pedestrian images captured by road cameras,and then uses the modified YOLOv5 s network to train and test them for target detection.This algorithm utilizes the advantages of YOLO algorithm in single stage detection,and inputs the images to be detected to output the prediction results.In addition,to address the issue of edge device deployment,this paper has also made lightweight improvements to the YOLOv5 algorithm,mainly through network sparsity,batch normalization(BN)layer optimization,key layer pruning,and algorithm fine-tuning methods,significantly reducing the model size of the final algorithm,to meet the needs of edge device deployment with small footprint and high accuracy.In the indoor scene,this paper uses two types of publicly available CSI signal datasets,and performs preprocessing methods such as data cleaning on the signal data.After that,it uses a customized deep learning network algorithm for training and testing.This algorithm uses convolutional neural networks for signal spatial feature extraction,and uses long-term and short-term memory networks for signal temporal feature extraction,Finally,feature fusion and attention mechanism are introduced to improve the robustness and generalization ability of the algorithm.The research results show that the modified YOLOv5 s algorithm proposed in this paper is significantly superior to similar algorithms in pedestrian detection,with 88.65% m AP performance,and after lightweight,the memory footprint is only one eighth of the original model,significantly reducing the memory footprint,making it more suitable for deployment of edge devices.The Wi Fi passive detection method based on spatiotemporal feature fusion proposed in this paper achieves a recognition accuracy of over 97%,and also has significant advantages compared to other network algorithms. |
PDF Full Text Request