International Journal of Remote Sensing · 2006 · 41 citations · 13 references
ComputingComputer ScienceGeoinformaticsComputer VisionPavement EngineeringRemote SensingChange DetectionDigital Image ProcessingImage SegmentationRemote Sensing Image AnalysisRemote Sensing ImageryTexture AnalysisGrey Value AnalysisHigh-resolution Satellite Imagery AnalysisPixel-level Change DetectionProbability Map EstimationProbability MapsSatellite Image ClassificationTexture-based SegmentationVehicle Detection
We propose and investigate three methods for deriving vehicle counts and classes from 1‐m resolution satellite and airborne imagery. We aim to characterize the performance of the proposed methods under varying conditions of lighting, atmospheric conditions, pavement characteristics, and vehicle density. For this purpose a representative set of images was extracted from available imagery and results from the three algorithms evaluated. The first algorithm uses principal component analysis (PCA) to re‐orient a set of texture bands in a way that the separation between the vehicle class and the pavement class is maximized in one of the principal component bands. A suitable threshold is selected to segment this PCA band into two classes—vehicles and pavement. The second method, Bayesian background transformation (BBT), attempts to obtain a probability map that assigns a posterior probability measure to each pixel based on the change in grey value between the current image and the estimated background image. A suitable probability threshold is selected to segment vehicles from the pavement. The third method, the gradient based method, attempts to exploit the local change in grey values to arrive at the final segmentation. The method uses a first order gradient to characterize local change and then uses an automatic process to select the threshold for the gradient image. After the segmentation step, various geometric properties are used to sieve out spurious responses and cluster pixels into vehicle and to arrive at the final vehicle classes with corresponding counts. The methods were compared on a test suite of satellite and airborne imagery. The BBT method seemed to provide the best and most robust performance with low errors of omission and commission and accurate vehicle counts, while the other two methods gave mixed performances.
13
A System for Video Surveillance and Monitoring
Robert T. Collins, Alan J. Lipton, Takeo Kanade et al. · 2000 · 1.2K citations
Moving target classification and tracking from real-time video
Alan J. Lipton, Hironobu Fujiyoshi, R. Patil · 2002 · 1.1K citations
Nikhil R. Pal, Sankar K. Pal · Signal Processing · 1989 · 280 citations
Car detection in low resolution aerial image
Tao Zhao, R. Nevatia · 2002 · 159 citations