基于运动上下文学习的老鼠社会行为检测与识别研究

项目名称： 基于运动上下文学习的老鼠社会行为检测与识别研究

项目编号： No.61300111

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 自动化技术、计算机技术

项目作者： 张盛平

作者单位： 哈尔滨工业大学

项目金额： 28万元

中文摘要： 神经科学研究的一个关键手段是通过老鼠实验理解特定行为在大脑中的编码机理。人为观察老鼠行为耗时耗力且极易引入主观误差，不利于进行大规模实验。本项目研究利用计算机视觉技术自动地观察老鼠行为。与现有方法依赖视觉跟踪得到单只或两只老鼠的位置信息，继而识别其简单行为不同，本项目针对复杂场景下多老鼠的低区分性和严重遮挡导致的跟踪困难，研究基于运动上下文学习的老鼠社会行为检测与识别：1）模拟大脑初/中级视皮层时序信号的认知机理，基于无监督深层稀疏编码自底向上学习不随形状变化的运动特征；2）结合自动上下文和AdaBoost思想，学习由运动特征判别老鼠社会行为的运动上下文模型；3）为达到实时性，进一步研究利用高效的加速逼近梯度法和交替方向法求解稀疏编码系数和在线学习字典。本项目的研究，为神经科学家进行大规模老鼠实验提供有利的辅助工具，也为视频高级语义行为理解提供一套新的运动特征学习和分类框架。

中文关键词： 老鼠社会行为检测与识别；稀疏编码；深层特征学习；在线学习；运动上下文模型

英文摘要： A critical approach of neuroscience research is to understand how specific behavior is encoded in the brain by conducting mouse experiments. Observing mouse behaviors by a human is time-consuming, labor-intensive and vulnerable to the introduction of subjective errors, therefore is not good for conducting large-scale experiments. This project studies an automatic approach for mouse behavior observation by exploiting computer vision techniques.Existing methods rely on visual tracking to obtain position information of one mouse or two mice and then recognize some simple mouse behaviors. However, in a complex scene of multiple mice, the low contrast and severe occlusion between mice will cause some difficulties for visual tracking.To overcome these difficulties, we study mouse social behavior detection and recognition based on motion context learning: 1) By simulating early and intermediate levels of visual cortex, we propose a bottom-up motion feature learning method based on unsupervised deep sparse coding. The learned motion features are invariant to shape changes. 2) With auto-context and Adaboost, we automatically learn motion context models, which are capable of discriminating mouse social behaviors. 3) To achieve real-time processing, we study to exploit the accelerated proximal gradient method and the alter

英文关键词： Mouse social behavior detection and recognition；sparse coding；deep feature learning；online learning；Motion context model