Gait recognition with event cameras

Preview

PDF WRAP_Theses_Du_2021.pdf - Submitted Version - Requires a PDF viewer. Download (14MB) | Preview

Request Changes to record.

Abstract

Gait recognition is a fundamental task in activity tracking, health monitoring, security surveillance and many other computer vision applications. A variety of sensors have been utilised for gait recognition, such as standard cameras, infrared cameras, floor sensors and inertial sensors. However, each kind of sensor has its limitation by nature. Event camera is a new bio-inspired vision sensor with lower energy consumption, broad dynamic range, and high temporal resolution. These advantages enable event cameras to be suitable for surveillance tasks, especially under special conditions such as long-term, sensitive, and challenging lighting scenarios. Unfortunately, to the best of our knowledge, there has been no event-based gait recognition technique available before. In this thesis, we focus on enabling approaches and solutions on gait recognition with event cameras. Firstly, due to the lack of relevant data, we produce two event-based gait datasets using an event camera, which serve as a basis for model training as well as quantitative evaluations and comparisons. Secondly, we propose a CNN-based approach named EV-Gait which achieves gait recognition with event cameras, and devise a scheme that includes image-like representation, noise cancellation and a neural network. Thirdly, we further propose a GCN-based 3DGraph-Gait approach that extracts spatiotemporal features from event streams, which improves the accuracy of recognition, and enables real-time gait recognition that only requires a limited number of events generated in several milliseconds. Finally, since privacy is a major concern with gait recognition, we propose an encryption framework named EV-Encryp, which effectively protects personal privacy and meanwhile, preserves the efficiency of the follow-up gait recognition after decryption. In summary, this study has initialised a novel research direction namely gait recognition with event cameras, contributed innovative supporting techniques and solutions, and established key foundations for further exploration and extensions.

Request changes or add full text files to a record

Repository staff actions (login required)

View Item