模拟视皮层信息加工反馈机制的人体动作识别研究

Due to the wide application of human action recognition in many fields, such as anomaly detection and autonomous vehicles, it becomes one of the research focuses. In recent years, although many algorithms and methods have been developed, there are still many scientific problems to be solved.. Therefore, this project will model visual feedback mechanism of information processing of both visual cortical areas V1 and MT in the dorsal stream on the basis of previous research, develop a computational model of spatiotemporal information processing with feedforward and feedback connections. Firstly, based on center surround inhibition and facilitation caused by lateral connections (feedback connections) of neurons within a area, we study the operating functions of surround inhibition and facilitation (two opposite effects), and explore method of their mutual iterative operations and a decision function for equilibrium, and discuss convergence conditions. Secondly, based on the short feedback connection theory between two visual areas, we will focus on the feedback mapping method from MT to V1, and study on the relationship of feedback connections between the neurons in different areas, and discuss the strategies of feedback signals generating and processing. Finally, we further analyze the influence of the different properties of neurons in visual cortical areas on feature extraction from video,and explore new representation and recognition methods of human action. On this basis, we will establish a complete system for human action recognition. By testing on general videos and analysis of its performance, we will verify the effectiveness of new theories and new methods. Through the research of this project, it promotes the development of artificial intelligence and computer vision technology.

由于人体动作识别在异常事件检测、无人驾驶等领域的广泛应用，成为备受关注的研究热点之一。近年来，虽取得许多成果，但仍存在许多亟需解决的科学问题。为此，本项目拟在前期研究基础上，模拟视皮层背侧通路中初级视皮层区域(V1)和颞中区(MT)内、区域间信息加工反馈机制，构建具有反馈连接的时空信息处理计算模型；该模型模拟区域内神经元侧链接（反馈连接）所引起的环绕抑制和易化作用，研究两个作用（相反效果）的操作函数，探索它们相互迭代操作的方法，建立平衡判决函数，讨论平衡条件；同时，模拟区域间短距反馈连接，研究从MT至V1区域的反馈映射方法，建立映射关系，并探讨反馈信号的产生方法和控制策略；分析神经元属性对特征提取的影响，探索新的动作表达方法。在此基础上，建立完整的人体动作识别系统，并通过测试分析，验证新理论和新方法的有效性。本项目研究所提供的脑启发式动作识别理论与方法，将推动人工智能和计算机视觉的发展。

视觉感知作为场景分析和理解的主要手段之一，对复杂环境下人体行为识别发挥重要作用，成为备受关注的研究内容。目前虽已取得许多成果，但普遍缺乏鲁棒性。为此，本项目以此为载体，模拟视觉皮层信息处理机制，根据视觉皮层区域内、区域间神经元之间反馈侧链接的基本神经生理属性，开展了时空特征提取和与行为识别识别计算模型的研究。首先，利用神经元的方向选择性、速度选择性和时空不可分离性，以及区域内神经元之间侧链接所引起的中心环绕作用，提出了空间特征提取的计算方法，构建了基于时空特征的神经网络模型，实现了人体动作识别；其次，利用区域间神经元之间的侧链接所引起的中心环绕调制作用，建立了区域间相互作用的映射关系，提出了空间特征提取的计算方法，建立了基于非经典感受调制的神经网络模型，实现了图像中轮廓检测；再次，针对具有反馈机制的神经网络计算模型，利用LMI技术，讨论了其稳定性，给出了网络指数收敛的条件；最后，利用已构建的计算网络模型，实现了视频中人体动作识别和医学图像序列分割系统，通过实验测试分析，从多个角度验证新理论和新方法的有效性。本项目研究所提供的脑启发式计算理论与方法，将推动人工智能和计算机视觉的发展。