Fig. 5
From: Dog facial landmarks detection and its applications for facial analysis
Video Anomaly Detection Pipeline (left). In each video frame, 46 facial landmarks are detected, normalised, and organised into time series. These series are then processed using a sliding window and an LSTM autoencoder model. The autoencoder reconstructs the output sequence and then compares it to the input sequence. An anomaly is detected if the error between the two sequences exceeds a certain threshold. LSTM Autoencoder Architecture (right). The encoder and decoder consist of two bidirectional LSTM layers with ReLU activation and a recurrent dropout.
