视频侦查中基于深度学习的人体行为识别技术研究

Human action recognition, one of the key technologies of video investigation, has aroused widespread interest by understanding the suspect’s action semantic characteristics and helping catch the key suspect points. Traditional dictionary learning methods depend on plenty of labeled samples. However, labeled samples are hard to collect in real world video investigation, existing methods employ deep-learned features from convolutional network to recognise actions, but ignore the probability distribution rules of feature space, and the global action semantics between sub actions and activities, thus the performance of classifiers will degrade when labeled information are limited. Therefore, this project presents a joint framework of human action recognition which is based on model representation: firstly, we obtain action representation based on linear optical flow by modeling human action sequences; secondly, in order to improve the ability of action representation, we use deep Convolutional Neural Network and bayesian statistics to learn high-level semantics pattern; at last, we employ graph-based semi-supervised method for action recognition, and try to train classifiers by leveraging both labeled and unlabeled data simultaneously. The performance of classifiers still can be improved when the labeled samples are scarce. Compared to the classical methods, this project try to increase the accuracy by five to ten percent, and improve the operational efficiency of intelligent surveillance system.

人体行为识别作为视频侦查的关键技术，能理解视频目标对象的行为语义特性，帮助甄别关键嫌疑要点，引起了广泛关注。传统字典学习方法依赖大量标注样本，而视频侦查中标注样本难以获取，且现有行为识别方法利用卷积网络的深度特征预测人体行为，忽略了动作特征空间的概率密度分布规律，以及子动作与行为活动之间的全局语义特性，导致标注信息不足时分类器性能急剧下降。为此，本课题提出一整套基于模型表示的人体行为识别方法。首先，对人体行为的动作序列进行建模，快速获得基于线性光流模型的动作特征表示；其次，采用深度卷积神经网络和贝叶斯统计方法，获取样本弱标注信息以提升行为语义的表达能力；最后，将基于嵌入流形图的半监督学习方法引入到行为识别中，使得有标注数据和未标注数据同时参与分类器训练，解决标注样本不足情况下分类器性能受限的问题。预期平均准确率较现有代表性方法提升5-10%，显著提升智能监控系统的工作效能。

在社会治安形势日趋严峻的背景下，如何利用人工智能技术做到视频侦查“快侦快破”已成为一个研究热点。人体行为识别是机器视觉甄别特定肢体动作的有力工具，但实际城市监控场景的异常行为样本稀少导致模型性能受限，且市面上缺乏监控视频人体动作识别的智能系统。因此，本项目主要致力于理论研究和应用研究2个方面的工作。.本项目取得的重要结果有：.1、在人体行为识别的理论研究方面，基于深度卷积神经网络构建了一整套的视频人体行为分析理论，通过提出在线学习和离线训练、骨架数据自监督对比学习、判别流形投影矩阵最优化等算法，分别在线性光流特征提取速度、人体骨架动作语义属性、半监督分类器模型性能上优于传统方法。项目组所提半监督动作识别算法的平均精度较当时代表性方法提高5-12.53%。.2、在人体行为识别的应用研究方面，基于多纤维残差网络和伪标注协同训练方法，通过对未标注训练样本提供弱标注信息，以增加网络模型分类效果；结合目标检测和多目标跟踪算法，提取多个行人目标的各自活动区域，在多目标跟踪队列中更新各目标的行为类别。项目组研制的视频人体行为分析系统平台，已在公安部第三研究所和北京市大兴区落地开展应用示范。