Detection and tracking using wireless sensor networks

Nadeem Ahmed, Yifei Dong, Tatiana Bokareva, Salil S. Kanhere, Sanjay Jha, Travis Bessell, Mark Rutten, Branko Ristić, Neil Gordon2007 · DOI

TL;DR

This paper explores using inexpensive Wireless Sensor Networks (WSNs) for target detection and tracking, overcoming the limitations of costly, specialized sensors. They propose a Particle Filter for Track-Before-Detect (TBD-PF) approach that effectively estimates target presence, position, and intensity using acoustic signals from off-the-shelf sensor nodes.

What Problem Does This Paper Address?

Traditional target detection and tracking systems often rely on expensive, high-performance sensors, limiting their scalability and widespread deployment. While Wireless Sensor Networks (WSNs) offer a cost-effective alternative, they introduce challenges like real-time decision-making, high-frequency data sampling, complex signal processing, and data fusion. This paper tackles the problem of achieving robust target detection and tracking using low-cost WSNs for ground surveillance, specifically addressing the difficulties of processing data from these resource-constrained devices.

Methods

The core of the proposed solution is a Particle Filter for Track-Before-Detect (TBD-PF) algorithm. Unlike simpler threshold-based methods that might discard valuable sensor data, the TBD-PF explicitly models target presence and absence. It can track targets with random movement patterns and handle non-Gaussian noise in sensor readings, which is crucial for real-world deployments where sensor calibration errors and environmental noise are common. The filter also allows for detecting targets with varying intensity levels.

The system models particles as states including position, velocity, intensity, and presence/absence indicators. It uses a probabilistic approach to model target birth and death. For sensor measurements, the paper assumes a model where the measured acoustic signal is a combination of the target's contribution (attenuated by distance) and background noise. The system recursively estimates the posterior density of the target's state (presence, position, velocity, and intensity) using all available measurements.

Datasets

The paper describes an experimental setup using twenty-five Xbow MicaZ motes. These motes were programmed to sample acoustic signals at 5kHz generated by a remote-controlled toy car. The motes were deployed in a grid, and summary statistics (sum and square of sum) of the acoustic signals were transferred wirelessly to a base station for processing by the PF algorithm. This constitutes a real-world dataset collected from the experimental setup.

Key Insights and Ideas

  • Cost-Effective Tracking is Achievable: The paper demonstrates that robust target tracking can be achieved using inexpensive, off-the-shelf WSN devices, significantly lowering the barrier to entry for such systems.
  • Particle Filters Excel in WSNs: The TBD-PF approach proves to be a powerful tool for WSN-based tracking, offering advantages over simpler threshold-based methods by explicitly modeling target dynamics and handling uncertainties.
  • Modeling Target Presence/Absence is Crucial: Explicitly modeling whether a target is present or not within the particle filter framework is key to distinguishing true detections from noise and false alarms.
  • Real-World Noise is Manageable: The ability of the PF to incorporate non-Gaussian noise, derived from calibration, is vital for practical applications where ideal noise models don't hold.

Key Results

The experimental results presented are promising. Figure 4 shows actual and estimated target paths in black and red, respectively. These results demonstrate that the PF-based estimator performs "very well in tracking different trajectories of the target using real data collected by off-the-shelf WSN devices." While specific quantitative metrics like accuracy percentages or error bounds are not explicitly detailed, the visual evidence suggests successful tracking of the toy car's movement.

Limitations and Future Work

The paper acknowledges that the initial values of target intensity and position are unknown and are recursively estimated. It also assumes that the probabilities of target birth and death are known a priori; if not, a very low value is assumed. The paper does not delve into the computational complexity or energy consumption trade-offs of the TBD-PF algorithm on the resource-constrained MicaZ motes, which is a common limitation in WSN research. Future work could explore optimizing the algorithm for these constraints and investigating its performance with more complex target behaviors and environmental conditions.

Summary updated

Abstract

Target detection and tracking is a well-established area of research. However, a majority of proposed solutions in existing literature rely on expensive and specialized sensors, which often have limited coverage. Using low cost sensor nodes is an attractive and complementary approach to scalable target detection and tracking applications. However, tracking with low cost Wireless Sensor Network (WSN), presents its own challenges, namely real time decision making, high frequency sampling, multi-modal sensing, complex signal processing, and data fusion. In this work, we investigate the use of inexpensive off-the-shelf WSN devices for ground surveillance. Our system estimates and tracks a target based on the spatial differences of the target object's signal strength detected by the monitoring sensors at different locations.
BibTeX
@misc{ahmed2007detection,
  author = {Nadeem Ahmed and Yifei Dong and Tatiana Bokareva and Salil S. Kanhere and Sanjay Jha and Travis Bessell and Mark Rutten and Branko Ristić and Neil Gordon},
  title = {{Detection and tracking using wireless sensor networks}},
  year = {2007},
  doi = {10.1145/1322263.1322328},
}
Figure