TL;DR
This paper introduces a deep learning approach to count people in a room using Ultra-Wideband (UWB) Channel Impulse Response (CIR) signals. By exploring various neural network architectures and ensembling techniques, the researchers achieved high accuracy in estimating occupancy, demonstrating UWB's potential for non-intrusive people counting.
What Problem Does This Paper Address?
Accurately counting people indoors is crucial for various applications, from building management and resource allocation to enhancing safety in public spaces. Traditional methods like imaging technology face challenges with poor lighting, obstructions, and privacy concerns. Radar systems can be expensive, and other wireless technologies like Bluetooth and WiFi struggle with precise counting. Ultra-Wideband (UWB) technology offers a promising alternative due to its high frequency bands, wide bandwidth, and resistance to interference, enabling precise distance measurements. While UWB has been explored for localization, its application in people counting, particularly using deep learning on its Channel Impulse Response (CIR) signals, remained an area ripe for investigation. This paper aims to bridge this gap by developing and evaluating deep learning models for UWB CIR-based people counting.
Methods
The core of this research lies in applying deep neural networks to UWB CIR signals for people counting. The authors explored several neural network architectures, including:
- Fully Connected (FC) Networks: Investigated varying depths (1-3 hidden layers) and parameter counts.
- Convolutional Neural Networks (CNNs): Explored different parameter values and kernel sizes within convolutional layers.
- Mix Neural Networks: Combined CNN layers with FC layers.
- AlexNet Style Network: Adapted the AlexNet architecture for 1D CIR data.
- ResNet Style Network: Leveraged residual connections to train deeper networks.
Beyond individual models, the paper extensively investigated model ensembling (combining predictions from multiple trained models) and signal ensembling (merging multiple CIR signals before processing). They also explored hyperparameter tuning, including normalization layers (Batch and Layer Normalization), dropout, pooling (Max and Average), and optimization techniques like the ADAM optimizer. The loss function used was Cross-Entropy Loss.
Datasets
The authors collected their own dataset for this study.
- UWB CIR Dataset: Collected using Qorvo DWM3000 UWB modules in a controlled laboratory environment (12m x 8m x 2.6m). The experiment involved varying the number of people (0 to 5) in a Line of Sight (LoS) scenario. Data was acquired at a rate of two instances per second, with 500 CIRs measured per setting. The dataset includes variations in subject demographics (height, body shape, gender) and movement (static vs. moving). A total of 33,000 CIRs were collected across 11 different settings (4 static, 7 moving).
Key Insights and Ideas
- CIR Variance as a Proxy for Occupancy: The study demonstrates that the variance in UWB CIR waveforms increases with the number of people present. This is attributed to the additional reflected waves from multiple individuals.
- Motion's Limited Impact on CIR: Surprisingly, the motion of individuals had a less significant impact on the CIR compared to the number of people. The authors attribute this to the high frequency of UWB signals.
- Deep Learning Outperforms Traditional Methods: Deep neural networks significantly outperformed traditional methods like Decision Trees, Naive Bayes, and K-NN classifiers, especially when dealing with raw or PCA-reduced CIR data.
- Ensembling is Key for High Accuracy: While individual models showed promise, both model and signal ensembling were crucial for achieving the highest accuracy. Combining multiple signals, particularly, showed a substantial performance boost.
- Strategic Ensembling for Efficiency: The paper highlights that a strategic combination of a few models and signals can achieve near-benchmark performance with significantly fewer computational resources (FLOPs) compared to using a larger number of signals or models.
Key Results
- Single CIR Accuracy: The best-performing single model achieved an accuracy of 81.86% using a single CIR.
- Model Ensembling: Ensembling models resulted in an approximate 2-3% performance improvement.
- Signal Ensembling: Combining multiple CIR signals led to significant accuracy gains. Using just four CIR signals with an ensemble of FC Net and ResNet models achieved 99% accuracy.
- Efficiency Trade-off: The paper demonstrates that achieving high accuracy (e.g., 99%) with four signals and two models was more computationally efficient (less than half the FLOPs) than using ten signals.
- ResNet's Robustness: The ResNet model showed superior performance in highly variable signal conditions, indicating its strength in handling challenging environments.
Limitations and Future Work
- Limited Number of People: The experiments were conducted with a maximum of five people. Future work aims to expand the system for estimating larger numbers of people, which is relevant for applications like maintaining room occupancy limits.
- Environmental Dependence: The study acknowledges that CIR waveforms are significantly affected by the environment (e.g., presence of furniture like desks). This implies that training models for each specific environmental condition might be necessary.
- Moving Objects: The influence of moving objects within the environment can further complicate CIR analysis. Future work should include training data that accounts for object motion.
- Multiple Rx/Tx: The current setup uses a single Rx and Tx. Future research plans to explore systems with multiple Rx and Tx units to enhance estimation capabilities.
Summary updated