Video Action Recognition

Action Recognition

ONNX

CNN+LSTM

This application implements a hybrid action-recognition architecture that combines a 2D Convolutional Neural Network (CNN) for spatial feature extraction with an LSTM-based recurrent module for temporal modeling. The model processes short sequences of consecutive RGB frames (e.g., 8–32), extracting per-frame visual descriptors through a lightweight CNN backbone. These descriptors are then fed into a bidirectional or unidirectional LSTM layer that captures motion dynamics and temporal dependencies, enabling robust classification of human actions in real-world environments. The pipeline includes temporal sampling, frame normalization, optional optical-flow augmentation, and sequence-level aggregation. The ONNX-exported model is optimized for real-time edge deployment, ensuring low-latency inference on SECO hardware for workflow monitoring, safety applications, and activity understanding.