Instance-Wise Dynamic Sensor Selection for Human Activity Recognition

Xiaodong Yang, Yiqiang Chen, Hanchao Yu, Yingwei Zhang, Wang Lu, Ruizhe Sun

Abstract

Wearable Human Activity Recognition (HAR) systems typically face a trade-off between recognition accuracy and battery life. This paper presents Instance-wise Dynamic Sensor Selection (IDSS), a framework that optimizes sensor usage in real-time. By modeling sensor selection as a Markov Decision Process (MDP) and employing a novel "Mutual DAgger" imitation learning algorithm, IDSS identifies the most relevant sensors for specific activities. Experimental results on MHEALTH, PAMAP2, and ActivityNet datasets demonstrate that IDSS significantly reduces the number of active sensors while maintaining high classification accuracy.

Introduction

Modern health-tracking wearables rely on multiple inertial sensors—such as accelerometers, gyroscopes, and magnetometers—to monitor human activities. However, continuous operation of all available sensors leads to rapid battery depletion and high computational overhead. Most current HAR systems utilize a fixed sensor configuration, which often results in redundant data collection for simple tasks or insufficient data for complex ones.

The challenge lies in developing an instance-wise approach that dynamically activates only the necessary sensors based on the current context. This paper introduces IDSS to address this need, providing a scalable solution for power-efficient activity recognition.

Methodology: The IDSS Framework

The IDSS framework treats sensor selection as a sequential decision-making problem, allowing the system to adapt its sensing strategy at every time step.

MDP Formulation

The selection process is modeled as a Markov Decision Process with a continuous state space. Unlike previous methods that discretize the state space and lose information, IDSS preserves the fluid nature of human movement. The state incorporates features from currently active sensors and the probability distribution of predicted activities. The agent's action involves selecting a specific subset of sensors to activate for the subsequent time step. The reward function balances classification accuracy against the cost of sensor activation, controlled by a penalty parameter $\lambda$.

Mutual DAgger Algorithm

A key contribution is the Mutual DAgger (Dataset Aggregation) algorithm. Training a dynamic selector presents a circular dependency: an effective classifier is needed to evaluate sensor utility, while a robust sensor policy is required to train the classifier on sparse data. Mutual DAgger solves this by iteratively optimizing both components. It leverages expert "hindsight" knowledge of true activity labels to guide the selection policy and uses the policy's sparse selections to augment the classifier's training set. This imitation learning approach also enables the policy to recover from suboptimal sensor choices, enhancing real-world robustness.

Experimental Evaluation

The researchers evaluated IDSS using three benchmark datasets:

  • MHEALTH: 12 activities using 7 sensors.
  • PAMAP2: 9 activities using 9 sensors.
  • ActivityNet: 4 basic activities using 5 sensors.

Key Results and Analysis

Sensor Reduction and Accuracy Trade-offs

IDSS demonstrates a superior ability to minimize sensor usage without compromising performance. For instance, in the MHEALTH dataset, the average number of active sensors was reduced from 7 to approximately 1.78 while maintaining high accuracy. Using a metric called the balance-index, the authors confirmed that IDSS achieves a more optimal trade-off between accuracy and cost compared to random selection or standard MDP approaches.

Figure 1: Impact of the penalty parameter $\lambda$ on the average number of active sensors across different datasets.

As shown in Figure 1, increasing the penalty $\lambda$ leads to a sharp initial decrease in the number of active sensors. The system intelligently drops the least informative sensors first, showing a graceful degradation of performance rather than a sharp drop-off in accuracy.

Instance-Wise Dynamic Selection

The core strength of IDSS is its ability to adapt to the specific "sensor signatures" of different activities.

Figure 5: Temporal visualization of sensor selection across MHEALTH, PAMAP2, and ActivityNet datasets. Black regions indicate active sensors, while the red dashed line represents the Activity ID.

Figure 5 illustrates the instance-wise nature of IDSS. It visualizes how the system dynamically switches sensors as activities change over time. For example, "Frontal elevation of arms" might require wrist sensors, while "Walking" might be identifiable with just an ankle sensor. This context-aware behavior ensures that energy is only expended when the data is truly informative.

Conclusion

IDSS provides a robust framework for extending the operational life of wearable HAR devices. By moving beyond fixed sensor sets to a dynamic, instance-wise selection policy, the system achieves significant power savings. Future work will explore the integration of diverse sensor modalities, such as heart rate monitors, and investigate the scalability of the policy architecture for high-density sensor networks.

Summary updated

Abstract

Human Activity Recognition (HAR) is an important application of smart wearable/mobile systems for many human-centric problems such as healthcare. The multi-sensor synchronous measurement has shown better performance for HAR than a single sensor. However, the multi-sensor setting increases the costs of data transmission, computation and energy. Therefore, the efficient sensor selection to balance recognition accuracy and sensor cost is the critical challenge. In this paper, we propose an Instance-wise Dynamic Sensor Selection (IDSS) method for HAR. Firstly, we formalize this problem as minimizing both activity classification loss and sensor number by dynamically selecting a sparse subset for each instance. Then, IDSS solves the above minimization problem via Markov Decision Process whose policy for sensor selection is learned by exploiting the instance-wise states using Imitation Learning. In order to optimize the parameters of the activity classification model and the sensor selection policy, an algorithm named Mutual DAgger is proposed to alternatively enhance their learning process. To evaluate the performance of IDSS, we conduct experiments on three real-world HAR datasets. The experimental results show that IDSS can effectively reduce the overall sensor number without losing accuracy and outperforms the state-of-the-art methods regarding the combined measurement of accuracy and sensor number.
BibTeX
@inproceedings{yang2020instance,
  author = {Xiaodong Yang and Yiqiang Chen and Hanchao Yu and Yingwei Zhang and Wang Lu and Ruizhe Sun},
  title = {{Instance-Wise Dynamic Sensor Selection for Human Activity Recognition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i01.5461},
}
Figure