2010 · 38 citations · 6 references
ComputingComputer ScienceElectrical EngineeringFeature EngineeringMachine LearningMechatronicsAutomotive SensorsClassificationFeature DetectionFeature ExtractionMagnetic SensorsPattern RecognitionComputational Feature ExtractionFeature Engineering For Sensor DataLow-power Magnetic SensorsMagnetic Field SensorsPeak Detection FeaturesSignal ClassificationSignal Processing For Sensor DataTransceiver Design
We extend our work in vehicle classification. Our system is based on a low complexity wireless sensor network. The system consists of a low power microprocessor together with AMR magnetic sensors and an RF transceiver. Two AMR magnetic sensors are employed to extracts dominant low-complexity features including vehicle count, speed, length, Hill-pattern peaks, and normalized energy. These features yield a promising result when vehicle classification is based on sizes (96%). However, when classification of similar sizes, e.g. cars, vans, pickup trucks are studied. The results are relatively lower at 77%. The contribution of this paper include (1) the implementation of feature extraction (count, speed, length) on sensor board and (2) the study for additional different low-complexity features such that better classification rate of small vehicles is obtained. These features include Hill-pattern peaks and magnetic signal differential energy normalized to the vehicle speed and length. This paper proposed vehicle classification tree based on above extraction features. Our work focuses on low computational feature extraction and classification processes suitable for implementing on micro-controller. The same data set employed in is analyzed. The classification rate yield 100 percent for motorcycle, 82.46 percent for car, 78.57 percent for van and 65.71 percent for pickup. The overall accuracy is 81.69 percent.
6