Behavior Recognition of a Broiler Chicken using Long Short-Term Memory with Convolution Neural Networks

This study proposes a scheme for identifying broiler behavior using object detection and recurrent-based artificial neural networks. First, the trained YoLov4 object detection model was utilized to identify five pre-defined parts of the broiler chicken, and then, the chicken skeleton was constructed from these parts and the angle between the backbone fulcrum vectors was extracted. Finally, six broiler behaviors were detected through a time-series-based long short-term memory (LSTM) network. The chicken behavior recognition scheme has been validated in an outdoor environment. The average precision, average recall and F1-score obtained by the proposed scheme were 82%, 81% and 81%, respectively. The performance comparison compared to using the multilayer perceptron (MLP) network with a YoLov4 model was also discussed and analyzed in this paper.