Low-altitude Remote Sensing Image Detection And Tracking

Low-altitude remote sensing images in real life is becoming more and moreimportant, no matter when both large geological disasters (such as earthquakes) orlarge venues emergency situations (such as fire)happens, the use of low-altituderemote sensing image detection and tracking helps to timely access to the disaster areathe severity of the disaster and the spatial distribution, more conducive to betterdetermine the first time an emergency program that will reduce the loss of life andproperty to a minimum.Aiming at pixels which is in low-altitude remote sensing image recognition andtracking that appear in the body is very small percentage (approximately only about4-8pixels), what we want to realize human identification becomes very difficultsituation considering the traditional body characteristics (such as human faces, humanbehavior), and other features, so we use of low-altitude remote sensing imagesHaar-like features with a human shadow conducted to identify the body.In this paper, we find three-point to lower the rate of sample error through usingadaboost training devices were detected in the low-altitude remote sensing images.The first method, based on bootstrapping is forcused on training adaboostnegative sample selection; Firstly, considering the background image with the targetobject, selecting a target area of interest to set as a detection window, then, detectobject with an interception positive samples, after through the detection of the firstphase of the stage, there will be a lot of training to generate a lot of adaboost detectionwindow, if the window is not detected target interception, then it will be added to thenegative training adaboost samples, repeat this process will eventually get a bignegative sample library; this method can reduce the error detection rate adaboosttrainer.The second method, the final confidence value of Adaboost training device isobtained by accumulating the value weighted by the respective stages of the weakclassifiers. Friedman article pointed out that if positive samples and negative samplesis not mixed, adaboost trainer does not sensitive to the boundary value, so in order toreduce the error rate of detection which is consist of human form of the background image, through the final training adaboost confidence divides the number of weakclassifiers to get the value of the standard test, each negative sample interception, nota random choice, but the choice of standard detection value is greater than zero, thenegative samples close to zero, adding negative sample library. This strategy is moreconcerned about the confusing boundary sample and the process of enhancing thetraining of the border, helps to further reduce the error detection rate.The third approach in which each positive sample contains some backgroundimages, though sometimes indistinguishable to the naked eye, but when you pass thefeature to adaboost trainer training, negative sample information, which ischaracterized by signs of positive samples will generate a negative impact onextracting, so through the training process we choose select negative samples nearpositive samples to be added to the negative sample training, as a result it willstrengthen the negative sample information helps to reduce the training processadaboost error detection rates.Validated by experimental data comparison chart of the proposed three methodsin remote sensing image detection,adaboost training can largely reduce the errordetection rate.Low-altitude remote sensing image data obtained by the detection phase firstly isevaluated by Gaussian kernel function and treated by of non-maxima suppression,only confidence value exceeds a specified threshold was passed to the target trackingfunction, then use M-HMT optimal algorithm for target track treatment. Experimentalresults show that the optimal M-MHT algorithm can overcome missed, false alarms,etc. to achieve the right people in the low-altitude remote sensing images tracking.