其他知识库

针对视频序列中人体行为检测的问题,提出一种基于边界敏感网络的时序行为候选生成算法,在原有边界敏感网络的基础上通过对时序评估模块和候选评估模块引入更深层的卷积神经网络,进而对视频特征有更好的表达。同时在后处理阶段,在NMS(non-maximum suppression)算法中引入新的置信度分数高斯加权衰减方法。实验结果表明,该算法可以有效提高行为检测问题中时序行为候选生成任务的召回率。在公开数据集Activity Net上,提出的方法在保证生成相同数量候选的同时有更高的平均召回率。

Aiming at the problem of human action detection in video sequences，a temporal action proposal generation algorithm based on boundary sensitive network was proposed． Based on the original boundary sensitive network， a deeper convolutional neural network was introduced into the temporal evaluation module and the proposal evaluation module，thus to better express the characteristic of the video． At the same time，in the post-processing stage，nonmaximum suppression ( NMS) algorithm was improved by introducing a new confidence score Gaussian weighted attenuation method． The experimental results show that the proposed algorithm can effectively improve the recall of the temporal action proposal generation task in the human action detection problem． On the open dataset ActivityNet，the proposed method has a higher average recall rate while ensuring the same number of candidates are generated．