摘要
实验鼠行为分析数据是神经学、生理学、药物学等学科实验数据的重要部分。针对实验鼠缺少肢体运动信息的特点,提出一种实验鼠多行为分析方法。提取实验鼠轮廓的帧间变化信息,同时考虑变化信息与实验鼠本身的位置关系,对行为视频形成系列轮廓变化帧。在训练阶段,通过Pillar K-means聚类算法从系列帧中提取80个关键帧,并把每一个训练行为视频用对应关键帧频数的直方图表示。在测试阶段,测试视频用最近邻法确定每一帧对应的关键帧,形成相应的关键帧直方图,从而把分类问题变成一个直方图相似性问题,再应用卡方距离进行分类。实验结果表明,该方法对8种行为的分类准确率最高达到100%,最低达到95%。
Mice action data is an important part of the experimental data in neurology, physiological pharmacology, etc. A method is proposed on multi actions analysis of mice in this paper for lacking the limbs information. The inter-frame difference of mice silhouette are extracted while the relation between the difference and the position of mice silhouette is taken into consideration, then sequential silhouette difference frames are obtained from action videos. In training phase, the 80 key frames are extracted using Pillar K-means algorithm, each video is presented by the key frames and a histogram on frequency of key frame is obtained. In test phase, the histogram of each video is determined using its key frames by nearest neighbour algorithm. So, a classification problem is transformed into the similarities problem. Actions are classified by X2 distances. Experimental results show that the correct rate of the proposed method is a maximum of 100%, and the lowest of 95%.
出处
《计算机工程》
CAS
CSCD
2014年第3期213-217,223,共6页
Computer Engineering
基金
安徽省自然科学基金资助项目(KJ2011A251)