A Novel Wildlife Poaching Detection Solution using Spatio-temporal Data with Dynamic Time Warping
Anika Puri · 2021
Wildlife poaching of endangered species such as elephants and rhinoceroses in Africa and Asia for illegal trading has become a biodiversity crisis, which has also been recognized by a United Nations Sustainable Development Goal of halting biodiversity loss. The demand for ivory has decimated the elephant population. The main issue associated with attempts to tackle this crisis is the vast area of the wildlife national parks. Recently, unmanned aerial vehicles (UAVs) equipped with heat-sensing infrared cameras have been deployed to help park rangers survey large areas of national parks at night when > 80 % of poaching occurs. In order to maximize surveillance area (with fixed flight time and battery constraints) while avoiding detection, UAVs need to fly at high altitudes (>400ft). This results in very few pixels associated with the objects of interest, i.e., humans and animals, in these thermal infrared videos. Current state-of-art methods utilize shape-based object detection techniques to identify animals/humans in these thermal images, with detection accuracy of only 20 %. In order to battle this inaccuracy, this research presents a novel solution that exploits video data's spatio-temporal nature (differences in movement pattern of animals/humans over time, i.e. turning radius, speed) to derive unique time series. These spatio-temporal series are then classified as human or animal by leveraging dynamic time warping metrics with K-Nearest Neighbor Clustering. When tested on a real-life thermal infrared videos dataset (BIRDSAI), collected in partnership with four African national parks, this method was able to detect human activity with 94 % sensitivity enabling real-time inferencing. Furthermore, this solution has the potential to eliminate the need for high-resolution high-cost $4,800 nighttime-thermal cameras with commodity thermal cameras (<$250), as demonstrated by a design prototype.