Wireless Video Noise Classification for Micro Air Vehicles

Onboard processing of video is currently outside the capabilities of power limited micro air vehicles (MAVs), which forces researchers to transmit video and telemetry to a ground control station for capture and offline processing. Unfortunately, wireless video transmission can introduce structured noise, which can corrupt image processing algorithms if the noisy frames are not identified and rejected from further processing. In this paper, we describe the design and evaluation of a supervised learning based classifier for labeling frames of video ”noisy” or ”clean”. We pose the classification problem as one of texture classification in video, where texture is represented using a feature set including informative statistics of steerable pyramid coefficients and the principal components of the chromatic histogram. The main contribution of this paper is the definition of this feature set as determined from analysis of noisy video. We evaluate nine different binary classifiers using cross validation on a MAV video training set, with additional evaluation on two validation sets collected from different times, days and locations. Results show that the best performing classifier is stepwise logistic regression, with a cross-validation accuracy of 0.89.