摘要
提出一种能适应不同方向目标检测,可以有效减小训练样本集数量的目标检测算法。算法包含两部分:ISM(Implicit Shape Model)形状模型生成和目标检测。ISM形状模型中包含目标描述表和ISM形状模型两部分。目标检测时将图像中的局部特征与目标描述表进行匹配,同时结合ISM形状模型生成投票空间。通过在投票空间中搜索局部极大值,并采用自顶向下的分割和MDL算法来剔除虚假目标,获取图像中的目标检测结果。编程实现了该算法,并用汽车、摩托车、行人等典型目标进行了目标检测试验。试验结果证明该算法对复杂背景下目标检测有较好的性能。用含不同角度目标的图像与原算法进行了对比实验,实验结果表明提高了算法对目标角度变化的适应能力。
We present a target detection algorithm which is able to adapt to target detection with different directions and effectively cut down the number of training sample sets. The algorithm consists of two parts： implicit shape model（ ISM） generation and target detection. The implicit shape model contains target codebook and implicit shape model itself. In target detection procedure,local characteristics in test image are matched with target codebook,at the same time the implicit shape model is combined to generate voting space. By searching local maximum values in voting space,and applying the top-down segmentation and MDL algorithm to weed out the false targets,the target detection results are acquired from test image. We implement the algorithm proposed in the paper through programming,and carry out the target detection test by using typical targets of cars,motorcycles,pedestrians,etc. Test results indicated that this algorithm has good performance in detecting the targets with complex background. Contrast experiments for the images with different target angles and for the original algorithm are carried out,the experimental result indicates that the algorithm improves the ability in adapting to the target angle changes.
出处
《计算机应用与软件》
CSCD
北大核心
2014年第4期219-222,共4页
Computer Applications and Software
基金
航空科学基金项目(20120177004)
