High-Resolution Radar Sensors for Human Gait Classification

Abstract

Video cameras are widely used for gait-based surveillance systems. However, they raise privacy concerns in sensitive environments such as private homes or restricted areas. As a result, radar-based methods are being explored as a privacy-preserving alternative. These methods are particularly promising with the emergence of high-resolution radar sensors capable of operating effectively in indoor conditions.

This thesis investigates a classification pipeline that uses radar-based gait data to identify different walking patterns relevant to surveillance scenarios. The input data is preprocessed into radar RGB spectrograms of standardized size and fed into a Convolutional Neural Network (CNN) architecture. Three classes are considered: walking, walking with hands in pockets, and walking while carrying a box. Multiple CNN architectures were explored and optimized, including experiments with different input channels, convolutional depths, and pooling methods. The performance of the trained models is evaluated using separate training, validation, and test datasets.

The final model achieved high validation accuracy but showed a drop in test performance, suggesting signs of overfitting. Results indicate that while CNN-based classification is feasible for real-world gait analysis from radar data, careful attention must be paid to model complexity and dataset quality to improve generalizability.

Popular Abstract

In this thesis, a radar-based gait classification system was developed to distinguish subtle arm movements between different walking patterns relevant to surveillance and security. Using machine learning techniques and range-Doppler signatures, the system identifies motion behaviors such as walking, carrying, and concealing actions. Exploring the challenges of real-world gait recognition, from data collection to classification accuracy, and proposes an approach that is both robust and adaptable to varied operational environments.

The work involved collecting radar data using a synchronized camera-radar setup to capture real human motion sequences. Each frame will then be processed into range-Doppler images—frequency-distance representations that encode the dynamic features of body movement. Convolutional neural network (CNN) architectures were designed and trained on these images in a sliding-window fashion to automatically extract features and classify the gait type. The networks were tuned using different preprocessing strategies, architectural variants, and performance metrics to achieve robust classification accuracy.

One of the most interesting findings was the sensitivity of classification performance to noisy transitional frames. After removing these and optimizing the CNN structure, the model achieved improved reliability and interpretability. Suggesting that, when processed correctly, the radar data carry enough distinctive features to differentiate these slight variations between gait types. Results also show that the CNN models performed well with acceptable accuracy given the small amount of self-collected data in limited time.

The thesis addresses a growing need in modern surveillance systems: the ability to detect and respond to specific human behaviors without relying on cameras, which often raise privacy concerns. Traditional vision-based methods, while powerful, may be unsuitable in low-light or occluded environments and can be intrusive. Radar-based approaches, in contrast, are anonymous, robust to lighting conditions, and can operate unobtrusively in both indoor and outdoor settings. By focusing on behavior rather than identity, this method supports ethical monitoring while preserving individual privacy.

The relevance of this work lies in its potential to enable intelligent, context-aware surveillance systems that can flag unusual activity without human oversight. In environments such as airports, border checkpoints, or secure facilities, early detection of abnormal motion—such as concealed carrying or hesitant movement—can enhance response times and prevent incidents before they escalate.

Beyond security, the system can be adapted for healthcare monitoring, such as mobility issues in elderly patients, where privacy and non-intrusiveness are also key. The modular design of the pipeline allows for future extension to more complex actions or integration with other sensors.

Overall, this thesis contributes to the growing field of radar-based human activity recognition and presents a framework that balances technical feasibility with ethical awareness.

• author: Wan, Kuan Teh and Habib, Khadijah
• supervisor: Daniel Sjöberg
• organization: Department of Electrical and Information Technology
• course: EITM02 20251
• year: 2025
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• keywords: FMCW radar, Gait classification, CNN, Machine Learning, Deep Learning
• report number: LU/LTH-EIT 2025-1072
• language: English
• id: 9201704
• date added to LUP: 2025-06-18 08:38:52
• date last changed: 2025-06-18 08:38:52