Introduction
Traditional human identification systems rely heavily on visual cameras, which face limitations in low-light environments and pose significant privacy risks. Radar technology offers a compelling alternative, providing robust sensing capabilities that are inherently privacy-preserving. However, most existing Radar Human Identification (RHI) research focuses on "trimmed" data—short segments containing a single individual. In real-world scenarios, radar sensors capture continuous, "untrimmed" streams. This paper introduces the Radar Human Localization (RHL) task, which requires identifying both the identity and the temporal boundaries of a specific person within these continuous streams.
The Radar Human Localization (RHL) Task
The RHL task moves beyond simple classification to precise temporal pinpointing. Given an untrimmed micro-Doppler (MD) signature and a target identity, the system must determine exactly when that individual appears and disappears in the recording.
Understanding Micro-Doppler Signatures
The core data used for this task is the micro-Doppler signature. These signatures capture the unique frequency shifts caused by the movement of various body parts. For instance, the rapid swing of arms and legs creates distinct patterns compared to the slower, more rhythmic movement of the torso. These patterns serve as a "gait fingerprint" for identification.
The Causal Localization Network (CLNet)
To address the RHL task, the authors propose the Causal Localization Network (CLNet), which utilizes three specialized modules:
- Patch-wise Doppler Encoding (PDE): This module uses a sliding window and CNN to extract features from MD signatures. A window size of approximately 3 seconds (45 frames) was found to be the "sweet spot" for capturing a full human gait cycle.
- Target-guided Doppler Encoding (TDE): Since the system is looking for a specific person, it employs a Transformer-based attention mechanism to focus on signal components that match the target's gait characteristics.
- Causal Localization Head (CLH): This module implements a sequential prediction strategy. It first predicts frame-level scores, then the start time, and finally uses both to determine the end time. This causal flow narrows the search space and stabilizes the training process.
Dataset: IDRad-TBA
Due to the lack of existing untrimmed radar datasets for localization, the researchers developed IDRad-TBA (Temporal Boundary Annotation). Built from the IDRad dataset, it contains 5,000 synthesized MD signatures. These were created by stitching together segments of real recordings using a gradual region-based method to ensure realistic transitions and prevent the model from learning artificial "seams."
Results and Key Findings
CLNet demonstrates a significant performance leap over traditional models adapted for localization. On the IDRad-TBA test set, CLNet achieved a Recall@1 (at IoU=0.7) of 13.20%, vastly outperforming a standard deep CNN baseline (2.31%). At a more relaxed threshold (IoU=0.3), the model reached 41.99% R@1, proving its effectiveness in locating targets within the data stream. Ablation studies confirmed that the causal prediction approach—predicting start time before end time—is vital for maintaining high localization precision.
Conclusion
The introduction of the RHL task and the CLNet architecture represents a significant step toward practical, privacy-preserving human monitoring using radar. While current work assumes a single person is visible at a time, future research will aim to extend these capabilities to multi-person environments and explore weakly-supervised learning techniques to reduce the need for expensive frame-by-frame temporal annotations.
Summary updated