ANIWATCH: CAMERA TRAP DATA PROCESSOR FOR DEEP LEARNING-BASED AUTOMATIC IDENTIFICATION OF WILDLIFE SPECIES

摘要

Camera trap equipment is mainly used to monitor the status of wildlife in protected areas. The existing data survey identifies wild animals through visual inteipretation. This process not only requires a long time, but also has the problem that the expertise of the investigator determines the reliability of the data. Recently, deep Learning in the field of image recognition has been detecting the object identification, object count, and the image description in the image with high accuracy. In this paper, we introduce the camera trap data processor (AniWatch) which can automatically database wildlife identification, animal count, and motion information by deep learning. To test the software performance, the Sobaeksan national park's Jukry ong eco-corridor was selected as a study area. First, we collected the camera trap data in the area. Since we need to detect moving objects in a fixed position, we performed data preprocessing through computer vision algorithms. Through the image tracking algorithm, the minimum bounding rectangle of the wild animal object was detected, and each frame was saved as an image. Because each image is of different size and resolution, we adjusted it to 100 x 100-pixel sizes to recognize it as training data. For deep learning, we applied a convolutional neural network (CNN) technique which is used in the image recognition field. Open source libraries (OpenCV. TensorFTow. and Keras) were used to implement the model, and the software was developed as a GUI application through Python. In the test results. AniWatch confirmed that it could reduce the time required for visual interpretation and minimize human errors. In the future, we will provide an automatic calculator of monitoring statistics by inputting camera trap data.