CAJ | 학술논문

重点研究视频序列中运动目标的提取、跟踪、分类等问题，最终实现了一个完整的目标检测系统。首先，考虑到算法的复杂度和检测的实时性需求，优选了背景差与帧差相结合的检测方法，通过背景实时更新，达到了一个较好的检测效果。其次，充分考虑了卡尔曼滤波跟踪的简单、高效等特性，通过计算当前目标与候选目标之间的欧氏距离来寻找最佳匹配目标和区分相交物体。再者，提出了一种先用Adaboost算法获得各个目标的初分类，再采用BP神经网络对目标进行细分类的方法，实践证明，该方法在保证对绝大部分人、自行车、汽车有很好的分类效果外，对大部分杂物也能有较好的区分。最后，通过Matlab编程，设计并实现了一个完整的运动目标检测系统。
중점연구시빈서렬중운동목표적제취、근종、분류등문제，최종실현료일개완정적목표검측계통。수선，고필도산법적복잡도화검측적실시성수구，우선료배경차여정차상결합적검측방법，통과배경실시경신，체도료일개교호적검측효과。기차，충분고필료잡이만려파근종적간단、고효등특성，통과계산당전목표여후선목표지간적구씨거리래심조최가필배목표화구분상교물체。재자，제출료일충선용Adaboost산법획득각개목표적초분류，재채용BP신경망락대목표진행세분류적방법，실천증명，해방법재보증대절대부분인、자행차、기차유흔호적분류효과외，대대부분잡물야능유교호적구분。최후，통과Matlab편정，설계병실현료일개완정적운동목표검측계통。
The extraction,tracking and classification of moving target in video sequence are studied,and a whole target detection system was achieved. Considering the complexity of the algorithm and the real?time requirement of the detection,the detection method combining background difference with frame difference is optimized and selected,and a better detection effect was achieved by updating the background in real?time. Since Kalman filtering tracking is simple and efficient,the Euclidean dis?tance between current target and candidate target is calculated to search for the best matching target and distinguish intersecting target. A novel method is proposed,which applies Adaboost algorithm to acquire the preliminary classification of each target, and makes use of BP neural network to classify the target in detail. The practical results prove that this method can ensure a preferable classification effect to majority of persons,cars and bikes,and distinguish most varia better. An integral detection sys?tem of moving target was designed and implemented by Matlab programming.