Human Activity Recognition Algorithm Based on Bidirectional Multi-Channel Feature Fusion

This paper designs a bidirectional spatiotemporal feature fusion algorithm for human activity recognition based on frequency modulated continuous wave radar. The algorithm takes the three-dimensional point cloud data of human activity collected by the radar as input, and adopts a dual channel feature extraction method in spatial feature extraction. The voxelated point cloud data is put into a convolutional neural network for extracting coarse-grained spatial information. At the same time, a multi-layer perceptron is used to extract fine-grained spatial information from individual points in the input point cloud data. Then, the coarse-grained and fine-grained spatial feature information extracted from two channels are fused as subsequent inputs. In terms of extracting temporal features, this algorithm uses Bi-LSTM to extract temporal feature, which can simultaneously link temporal information before and after for feature extraction. This network model can not only extract local and global features from point cloud data, but also consider the information before and after when extracting temporal information, which can greatly improve the accuracy of human activity recognition.