KLASS - Human Recognition

KLASS - Human Recognition

Team members

Zhang Peiyuan (ISTD), Chong Chee Kit (ISTD), Wang Chenyu (ISTD), Cao Sen (EPD), Cai Xuemeng (ISTD), Ivan Tandyajaya (ISTD), Xu Yongren (EPD)

Instructors:

Kwan Wei Lek, Liu Jun

Writing Instructors:

Grace Kong

Teaching Assistant:

Congjian Lin

Problem

With the rise in security threats in recent years, as well as the methods to circumvent existing personnel identification systems, the world is becoming more dangerous to live in. Our revolutionary gait recognition system can be used to analyse the shape of an individual's body and the unique way in which that body moves, and detect suspicious persons unobtrusively.

Hassle-Free Data Collection

Collecting dataset that is relevant to your context, be it in public walkways or private estates, can be traditionally cumbersome. With our data collection pipeline, creating your own personalised dataset has never been easier!

Gait Recognition

Gait recognition technology (GRT) can analyse the shape of an individual's body and the unique way in which that body moves when walking or running. GRT could be used to monitor, track and identify people by the shape of their bodies and how they move in protests.

Product Pipeline

Data Collection and Gait Recognition

Data Collection

Collect Raw

Footage

Crop Footage with

MMDetection

Annotate

Videos

Data Collection

Completed

Gait Profiles

Upload gait

footage

Gait Profiles

Created

Gait Recognition

Select

timeframe

Crop and

select POIs

Send for gait

recognition

Gait Recognition

Completed

Data Collection Pipeline

Have a specific environment that isn't covered by existing datasets? Collect your own data and train your own model within days, using our data collection and annotation pipeline!

Blazingly Fast

It takes only 5 minutes to search for a person in the recorded footage, send that footage for gait analysis, and receive the top 5 predictions.

State-of-the-Art Model Accuracy

We collected our use-case-specific test set to evaluate previous models and datasets published in the gait recognition literature. Then, We pre-trained the models, including GaitSet, GaitPart, and GaitGL on the CASIA-B, OUMVLP, or GREW datasets, and picked the one with the highest accuracy.

We then trained it on our collected training set to further boost the performance on the test set. We found out that GaitSet was pre-trained with GREW, hence we combined it with our training set to achieve the best performance.