No abstract

Dogs and cats tend to show their conditions and desires through their behaviors. In companion animal behavior recognition, behavior data obtained by attaching a wearable device or sensor to a dog’s body are mostly used. However, differences occur in the output values of the sensor when the dog moves violently. A tightly coupled RGB time tensor network (TRT-Net) is proposed that minimizes the loss of spatiotemporal information by reflecting the three components (x-, y-, and z-axes) of the skeleton sequences in the corresponding three channels (red, green, and blue) for the behavioral classification of dogs. This paper introduces the YouTube-C7B dataset consisting of dog behaviors in various environments. Based on a method that visualizes the Conv-layer filters in analyzable feature maps, we add reliability to the results derived by the model. We can identify the joint parts, i.e., those represented as rows of input images showing behaviors, learned by the proposed model mainly for making decisions. Finally, the performance of the proposed method is compared to those of the LSTM, GRU, and RNN models. The experimental results demonstrate that the proposed TRT-Net method classifies dog behaviors more effectively, with improved accuracy and F1 scores of 7.9% and 7.3% over conventional models.