光学学报, 2018, 38 (8): 0810001, 网络出版: 2018-09-06   

基于空-时域特征决策级融合的人体行为识别算法 下载: 843次

Human Action Recognition by Decision-Making Level Fusion Based on Spatial-Temporal Features
作者单位
长春理工大学光电工程学院, 吉林 长春 130022
摘要
提出一种基于空-时域特征决策级融合的人体行为识别算法。在空间域提取人体的形状上下文特征,用于同一时刻模板图像与测试图像的轮廓匹配;在时间域用变化的空间特征序列表征运动特征,联合稳健的空间特征进行有效的行为识别。识别阶段采用动态时间规划算法分别计算两种特征对于每种类别的后验概率,在决策级采用加权平均法对两种特征的后验概率进行融合,将最大概率对应的类别记为最终分类结果。针对动态时间规划算法提出一种基于椭圆边界约束的改进搜索策略,有效缩减最优路径的搜索空间,同时剔除视频中的噪声干扰。从计算复杂度和识别精度两方面对椭圆边界的约束性能进行分析,实验表明,椭圆边界约束性能优于平行四边形及菱形约束,并给出最佳边界尺寸范围。算法分别在Weizmann、KTH和UCF101行为数据集上进行测试,平均识别率分别优于93.2%、92.7%和81.2%,有效实现了室内智能监控系统的高效性及稳定性。
Abstract
A human action recognition algorithm is proposed based on the decision-making level fusion with spatial and temporal features. Shape context feature of human body is extracted to match the contours of template images and test images in the spatial domain, while the motion feature is described by a changing spatial feature sequence in the time domain. Then, the motion feature is combined with the robust spatial feature for effective human action recognition. At the recognition stage, the dynamic time warping is applied to calculate the posterior probabilities of two kinds of features for each class. The weighted-average method is used to fuse the two posterior probabilities at the decision-making level, and the corresponding class with the maximum probability is recorded as the final classification result. Aiming at the dynamic time warping algorithm, we propose an improved searching strategy based on the elliptic boundary constraint, which can effectively reduce the space for searching for the optimal path, while eliminate the noise interference in the video sequence. The constraint performance of elliptical boundary is analyzed from two aspects of computational complexity and recognition accuracy. Experimental results show that the performance of elliptical boundary constraint is better than that of the parallelogram and diamond boundary constraints, and the optimal boundary size range is given. Experimental results on Weizmann, KTH and UCF101 datasets demonstrate that the average recognition rate of the proposed method is higher than 93.2%, 92.7% and 81.2%, respectively, indicating that the proposed method can effectively obtain the efficiency and stability of indoor intelligent monitoring system.

李艳荻, 徐熙平. 基于空-时域特征决策级融合的人体行为识别算法[J]. 光学学报, 2018, 38(8): 0810001. Yandi Li, Xiping Xu. Human Action Recognition by Decision-Making Level Fusion Based on Spatial-Temporal Features[J]. Acta Optica Sinica, 2018, 38(8): 0810001.

本文已被 3 篇论文引用
被引统计数据来源于中国光学期刊网
引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!