ReconGait: Giant Panda Gait Recognition Based on Spatio-Temporal Feature Reconstruction

As an endangered species, protecting and identifying individual giant pandas has always been a focus. However, traditional methods based on feces and molecular biology have inherent limitations. With the progress of deep learning, image analysis methods can deliver good results, but they rely on high-quality frontal facial images, which are challenging to acquire in the wild. Furthermore, these approaches have neglected the fact that a substantial portion of data from the wild is predominantly in the form of video, and have inadequately exploited the temporal information. Therefore, we apply gait recognition to identify individual giant pandas. By examining the unique walking patterns of individuals in the video, gait recognition allows accurate identification from various angles, long ranges, and without subject cooperation. This makes it ideal to identify wild individual giant pandas, but it also introduces new challenges, such as quadrupeds having more complex body gestures compared to humans, and the generally lower image quality taken by cameras in the wild. To tackle issues in natural environments, we design ReconGait, a giant panda gait recognition model capable of aligning multi-view images and restoring low-quality regions effectively. Experiments shows state-of-the-art results, proving the effectiveness of using gait analysis for individual giant panda identification. Through our work, we strive to contribute significantly to the preservation of this rare species and set a precedent for the integration of gait recognition in wildlife conservation efforts.