Memory Efficient Video Collection and Extraction Technique for Augmented Triplet Network's Dataset

Wild Deer grazing in farmers' and villagers' fields has caused serious trouble for the residents and concerned authorities in Northern Hokkaido of Japan. Furthermore, the number of deer in northern Japan is increasing rapidly leading to increment in road accidents. Deer-car collisions each year kill millions of people around the world which is one of the serious challenges that needs to be handled through scientific study. Several countermeasures have been applied to avoid or minimize to reduce deer-car crashes and the damages caused by the white-tailed deer, but a solid solution is yet to be found. In this context, we proposed a highly effective solution to identify the individual wild deer and pinpoint the most troubles causing the group of deer using Augmented Triplet Network (ATN), a novel technique of computer vision. Initially, the deer are recognized and classified individually which will help us track the deer and know the behavior of the deer in depth. The aim of this study is to reduce the video data by automatically extracting only the data necessary to construct a dataset for individual identification in the ATN model. The extraction method proposed in this study reduced the size of videos by 81 percent on average. We were able to obtain an accuracy rate of over 99 percent on two different datasets by using the ATN model.