A PCA-based vehicle classification system in wireless sensor networks

We designed, implemented, and evaluated a compact signal processing system on a Berkeley mote for vehicle classification via acoustic characteristics. The main components of this system include a sound sampler, FFT converter, an energy-based adaptive CFAR detector, and a classifier. We chose principal component analysis (PCA) as the primary classifier because of its high accuracy, good scalability, and low resource requirements. This system enables a Berkeley mote to detect a passing vehicle and to decide its type. Only the result of the identification analysis is reported to sink nodes. With this proof-of-concept implementation, we explored limits of system design on a mote. Using a testbed based on this system, we compared classification quality, energy consumption, and latency of the PCA classifier with two typical classifiers in both centralized and distributed signal processing scenarios. Briefly, we found that the PCA yields better classification quality but has slightly lower energy efficiency than the others. By extensive simulations, we studied the dependency of PCA on signal quality and efficiency issues of signal processing in WSNs