GaitSADA: Self-Aligned Domain Adaptation for mmWave Gait Recognition

Ekkasit Pinyoanuntapong1, Ayman Ali1, Kalvik Jakkala1, Pu Wang1, Minwoo Lee1, Qucheng Peng2, Chen Chen2, Zhi Sun3,
1University of North Carolina at Charlotte, 2University of Central Florida 3Tsinghua University
*The dataset is under review to ensure the preservation of the identity and privacy of individuals contained within.


mmWave radar-based gait recognition is a novel user identification method that captures human gait biometrics from mmWave radar return signals. This technology offers privacy protection and is resilient to weather and lighting conditions. However, its generalization performance is yet unknown and limits its practical deployment. To address this problem, in this paper, a non-synthetic dataset is collected and analyzed to reveal the presence of spatial and temporal domain shifts in mmWave gait biometric data, which significantly impacts identification accuracy. To address this issue, a novel self-aligned domain adaptation method called GaitSADA is proposed. GaitSADA improves system generalization performance by using a two-stage semi-supervised model training approach. The first stage uses semi-supervised contrastive learning and the second stage uses semi-supervised consistency training with centroid alignment. Extensive experiments show that GaitSADA outperforms representative domain adaptation methods by an average of 15.41% in low data regimes.

Overall Architecture of GaitSADA.

Interpolate start reference image.

(a) Stage 1. Pretraining via semi-supervised contrastive learning to learn a compact gait representation from both source and domain data, which also implicitly mitigates the source-domain distribution shift. (b) Stage 2. fine-tuning via semi-supervised consistency training with centroid alignment. The fine-tuning stage aims to further improve model generalization performance by pseudo-labelling the target-domain samples, clustering together the samples belonging to the same class but from different domains, and pushing the class centroid close to the weight vector of each class. The encoders are sharing the weights.

Data Collection

We curate a non-synthetic dataset consisting of mmWave radar-based gait biometric data. This dataset allows one to study and improve the spatio-temporal generalization performance of a radar-based biometric identification system

We collected gait data from 10 volunteers between the ages of 18-35. Each subject’s data was collected in four different locations. The source location was a research space with cubicles, and the other three areas consisted of a server, conference, and an office room. By maintaining four distinct locations, we introduce SDS. In the source location, data was collected on 10 different days for each subject. 5 separate days of data was acquired for each of the three other locations, which are used as target domains. A participant can either walk towards the radar or walk away from the radar, each of which is counted as one walking instance and generates one spectrogram data sample. The data collection was limited to 100 data samples per person in the source location and 50 data samples per person in each of the target locations on any given day

Texas Instruments (TI) IWR1642EVM boost board interfaced with a DCA1000EVM board to collect the raw mmWave radar signal. The radar system consists of two transmitting antennas, four receiving antennas, and 120° view of the azimuth plane. The radar system supports up to a 4Ghz bandwidth operating on 77 GHz to 81 GHz. To configure FMCW wave parameters such as chirp width, repetition time, and chirp slope from our radar device, we use a Dell Latitude 7480 laptop with TI mmWave studio software as a control system.

Interpolate start reference image.

The walking dataset is collected from four distinct locations such as laboratory, conference room, server room, and office. We recruit ten volunteers to walk in his/her natural ways. Each participant either walks toward or away from the radar. Data is preprocessed into 50 samples of five different days at each location, a total of 2500 samples (10 subjects × 50 samples × 5 days), except for the laboratory location, we collect additional data for another five more days (total of ten different days) in order to perform temporal domain shift experiments for a different time in the same location. We preprocess the data into 256×256×1 spectrogram and standardized values between -1 and 1. Laboratory location data is used as source data and the other location data (i.e. conference room, server room, and office) is referred to as target data. For the spatial domain drift experiment, day 1 to day 3 samples are used as a training set for all source and target locations. The rest of the data is used for testing. For the temporal domain drift experiment, we employ 1 to 3 days of laboratory location as source data and the next consecutive days of laboratory location as target data. For example, considering the temporal 2-day case, the first and the second-day data of laboratory location is a source domain, the third and the fourth-day data is a target domain, and the fifth to the tenth-day data is utilized as a test set.

Interpolate start reference image.


If you find our work useful in your research, please consider citing:

      title={GaitSADA: Self-Aligned Domain Adaptation for mmWave Gait Recognition}, 
      author={Ekkasit Pinyoanuntapong and Ayman Ali and Kalvik Jakkala and Pu Wang and Minwoo Lee and Qucheng Peng and Chen Chen and Zhi Sun},