CNN + BiLSTM

User leaving is not a single-frame event. It is a sequence: sitting, standing up, turning, and moving away. cnn_blstm_pointcloud_training_en.ipynb uses CNN + BiLSTM for this type of data. The aligned Chinese version is cnn_blstm_pointcloud_training_zh.ipynb.

Frame-Level Features

The CNN reads one frame of points and extracts spatial features from the point distribution.

Temporal Modeling

The BiLSTM reads the frame features over time. It can use context from both directions inside a segment, which helps classify the whole action clip.

Typical input shape:

[batch, sequence_length, num_points, point_features]

The model output is a class prediction that can be mapped to a security action such as locking the screen or hiding sensitive content.