视频序列中的目标检测

摘要

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。%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.