Deer Behavior Recognition with Deep Learning

─ Discussion of the Analysis Method for Camera Trapping Movie Using Body Part Trajectories ─

Abstract

Camera trapping is a method for analyzing wild animal behaviors and for wildlife damage control and ecosystem studies. Although this method can provide detailed information about animal behaviors, it still requires manual data confirmation, thereby making its analysis labor intensive. To automatically analyze camera trapping data, we created “BoxFlow images” in this study. These images depict locations and trajectories of deer parts in movies. In addition, these images were input into a convolutional neural network, from which we obtained outputs of three categories, namely, “Moving”， “Eating”， and “Distance.” A cross-validation-like approach was employed for the classification evaluation. When deer parts were successfully detected, mean accuracies of the behavior classification were 0.789 for Moving, 0.862 for Eating, and 0.872 for Distance.