Research

Gait recognition is a biometric technology that identifies individuals from their walking patterns, which is advantageous for long-distance and non-intrusive authentication. 

Our work in this area focuses on developing robust gait recognition models under varying conditions  (e.g., view angles, speed, clothing, carried objects, etc.), enabling reliable recognition. 

Beyond identity recognition, gait contains demographic information such as age and gender. Age-related changes, like children’s head-to-body ratio growth or middle-aged stoop, appear in walking patterns, enabling gait-based age estimation. Gender can be inferred from differences in hip and shoulder movement, stride length, and overall body dynamics. 

Our work includes building gait datasets for this task, proposing benchmark methods, and developing approaches for single-frame age and gender estimation as well as human mesh-based methods. 

In real-world scenarios, pedestrians are often partially occluded by obstacles, other people, or environmental constraints, posing significant challenges for gait recognition. Occlusion can lead to incomplete gait silhouettes and missing features, which degrade recognition performance. 

Our research addresses these challenges using human mesh-based modeling to reconstruct the full body and mitigate the effects of occlusion. I also investigate silhouette completion and registration techniques to handle missing regions in gait contour images. Additionally, I explore gait recognition with fisheye cameras, which are mounted on low-height robots; when people walk close to the camera, only partially occluded views are captured, and mesh-based models are used to reconstruct the full gait for accurate recognition.

In real-world environments, multiple people often appear in the same scene, leading to overlapping silhouettes and interactions that make gait recognition challenging. Traditional approaches that first perform detection and segmentation for each person, followed by recognition, often suffer from errors due to inaccurate detection or segmentation, resulting in degraded performance.

To address this, we proposed an end-to-end framework that simultaneously estimates the gait features and positions of multiple individuals, avoiding the errors introduced by stepwise processing. This approach significantly improves recognition accuracy in multi-person scenarios and is more robust to occlusion and crowding effects.

Electroencephalography (EEG) has recently emerged as a novel biometric modality. Unlike external traits such as face or gait, EEG signals are extremely difficult to imitate, since they can only be continuously acquired from a live subject using specialized equipment. While EEG has been widely explored for tasks such as emotion classification, it also holds great potential for user authentication.

In this work, we focus on EEG-based identity recognition in more natural and unconstrained (“in the wild”) settings, moving beyond the traditional controlled environments. In particular, we explore scenarios where users are engaged in interactive conversations, which better reflect realistic usage contexts. This approach provides new insights into building practical and robust EEG-based authentication systems for real-world applications.