BiLSTM-based Individual Cattle Identification for Automated Precision Livestock Farming

Individual cattle identification plays an important role for automation in precision livestock management. Existing methods for cattle identification require radio frequency and visual ear tags, all of which are prone to loss or damage. In this work, we propose a deep learning-based framework to identify beef cattle using image sequences, unifying merits of both Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) network methods. A CNN (Inception-V3) was used to extract features from a video dataset taken of the rear-view of cattle, after which extracted features were fed to a BiLSTM layer to capture spatial-temporal information enabling the identification of each individual animal. A total of 363 rear-view videos of 50 cattle were collected for our dataset. The proposed method achieved 91% identification accuracy using a 30-frame video length, improving that of Inception-V3 use or LSTM. Additionally, increasing video sequence length to 30-frames enhanced identification performance. Our approach can use spatial-temporal features to identify cattle, and enables automated identification for precision livestock farming.