MM Wave Radar Human Identification
This research outlines the critical architecture and methodology needed to achieve individual human identification (identifying which person is present) using mmWave radar, based on the principles demonstrated by the TI Sensor-to-Cloud project.
1. Data and Feature Focus: Gait and Micro-Doppler
The system's ability to identify individuals is fundamentally tied to capturing their unique movement signature:
- Goal: Human Identification relies on Gait Recognition—analyzing the subtle, unique motion patterns (stride length, arm swing, velocity changes) that distinguish one person's walk from another's.
- Key Feature: The most valuable feature for identification is the Micro-Doppler (µD) Signature. This is the subtle frequency shift caused by smaller moving body parts (hands, feet, head) and is highly unique to the individual.
- Data Source: While you can use raw ADC data, most successful identification systems use Point Cloud Data sequences because they are sparse, smaller, and easier to process on edge devices.
2. Model Architecture and Processing Pipeline
Simple fully connected networks are generally insufficient for temporal recognition tasks like gait analysis. You need networks that can understand sequences over time.
| Stage | Data & Processing | Recommended Method/Model | Relevance to our Project |
|---|---|---|---|
| Data Acquisition | Raw data is processed on-chip (IWRL6432) to generate the point cloud. | IWRL6432 (using the on-chip Hardware Accelerator) | Ensure chirp parameters give good velocity resolution to capture micro-Doppler shifts. |
| Feature Extraction | Capturing the temporal changes in position, velocity, and Signal-to-Noise Ratio (SNR) for each tracked person. | PointNet or custom Global Embedding layer | Crucial for extracting useful spatial features from the sparse point cloud data frame-by-frame. |
| Classification | The model processes the sequence of extracted features over several frames to learn the subject's identity. | Bi-directional LSTM (Bi-LSTM) or Temporal CNN (TCNN) | Planned 1D CNN is a good next step, but pairing it with an LSTM is critical for sequence learning (gait over time). |
| Tracking | Must assign a unique ID to each detected object in a multi-person environment. | Kalman Filter (Extended/Converted) | Necessary to ensure the ML model is being fed a clean sequence of points belonging only to the specific person you are tracking. |
3. Practical Implementation Considerations
- Edge Processing (On-Chip ML): Many successful TI implementations demonstrate that the classification models (sometimes quantized to 8-bit integer format) can run on-chip on the integrated Arm Cortex-M4F core, which is ideal for IWRL6432.
- Privacy: A significant advantage of radar over camera systems is that it maintains human privacy, as it does not capture visual images.
- Accuracy vs. Scale: Identification accuracy drops significantly as the number of individuals being identified increases and as people begin to occlude each other (stand in front of one another).