基于全局时空编码网络的猴类动物行为识别 |
| |
作者姓名: | 孙 峥 张素才 马喜波 |
| |
作者单位: | 1. 中国科学院自动化研究所,北京 100190;
2. 中国科学院大学人工智能学院,北京 100049;3. 北京昭衍新药研究中心股份有限公司,北京 100176 |
| |
基金项目: | 国家自然科学基金项目(82090051,81871442);中国科学院青年创新促进会(Y201930) |
| |
摘 要: | 猴类动物行为的准确量化是临床前药物安全评价的一个基本目标。视频中猴类动物行为分析的一
个重要路径是使用目标的骨架序列信息,然而现有的大部分骨架行为识别方法通常在时间和空间维度分别提取
骨架序列的特征,忽略了骨架拓扑结构在时空维度的整体性。针对该问题,提出了一种基于全局时空编码网络
(GSTEN)的骨架行为识别方法。该方法在时空图卷积网络(ST-GCN)的基础上,并行插入全局标志生成器(GTG)
和全局时空编码器(GSTE)来提取时间和空间维度的全局特征。为了验证提出的 GSTEN 性能,在自建的猴类动
物行为识别数据集上开展实验。实验结果表明,该网络在基本不增加模型参数量的情况下,准确率指标达到
76.54%,相较于基准模型 ST-GCN 提升 6.79%。
|
关 键 词: | 行为识别 骨架序列 全局时空编码网络 猴类动物 药物安全评价 |
Monkey action recognition based on global spatiotemporal encode network |
| |
Authors: | SUN Zheng ZHANG Su-cai MA Xi-bo |
| |
Affiliation: | 1. CBSR&NLPR, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China;2. School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China;3. JOINN Laboratories (Beijing) Co., Ltd., Beijing 100176, China |
| |
Abstract: | Accurate quantification of caged monkeys’ behaviors is a primary goal for the preclinical drug safety
assessment. Skeleton information is important to the analysis on the behaviors of monkeys. However, most of the
current skeleton-based action recognition methods usually extract the features of the skeleton sequence in the spatial
and temporal dimensions, ignoring the integrity of the skeleton topology. To address this problem, we proposed a
skeleton action recognition method based on the global spatiotemporal encode network (GSTEN). Based on the spatial
temporal graph convolutional network (ST-GCN), the proposed method inserted global token generator (GTG) and
several global spatiotemporal encoders (GSTE) in parallel to extract the global features in the spatiotemporal
dimension. To verify the performance of the proposed method, we conducted experiments on a self-built monkey
action recognition dataset. The experimental results show that the proposed GSTEN could achieve an accuracy of 76.54% without increasing the number of model parameters, which was 6.79% higher than the baseline model
ST-CGN. |
| |
Keywords: | |
|
| 点击此处可从《》浏览原始摘要信息 |
|
点击此处可从《》下载全文 |