计算机视觉中的运动分析及应用--刘允才剖析.ppt

摄像头意外检测 — 恶意破坏监控摄像头使之不正常工作 — 遮挡摄像头 — 偏转摄像头 — 强光照射、切断电缆等 — 场景检测 — 视频纹理信息 DEMO 跨场景多摄像机关联跟踪 — 跨场景的视频人物跟踪 — 不同场景 — 不同光照条件 — 不同人物姿态 — 不同的观察视角 — 基于多种特征数据融合的人体目标识别 — 多场景拓扑关系的辅助决策 DEMO 节能减排：智能空调 空调的三个控制要素 风向 、风量、温度设置 风向: 检测人体的空间位置 风量: 检测人体离空调机的距离 温度设置: 人员识别 形态、步态、动作特征… 计算机视觉能够胜任. 智能空调人员识别框架图 智能空调送风控制系统框架 Demo 仪器设备：无标记运动捕捉仪 Support 4-16 cameras GPU Acceleration — 400 times speed-up Demo 器人外科手术: 全膝关节置换手术 Common surgery for full knee replacement Italy: 16000 surgeries in 1997 USA: 26000 surgeries in 1998 Japan: 75000 surgeries in 1998 Germany: 55000 surgeries in 1998 In World: a million surgeries every year China: total 2000 surgeries full knee replacement Accurate limb alignment Good parenchyma balance Equal flexing gaps Technical requirement Position Accuracy 6 mm Limb Alignment 5 degrees Surgery Robot System： WATO-I and WATO-II WATO-I WATO-II WATO-II Configuration： Demo 2 Demo 1 Thank You GP-based: genetic programming based * 状态先验（初值）、似然函数设计、优化算法是生成式算法的三个主要因素， * * GPR — Generative Method, RVM — Discriminative method; Feature: HOG, LBP Shape descriptor (Belongie Malik): A shape is represented by a discrete set of points sampled from the internal and external contours on the shape. The shape context at a given point on a shape is invariant under translation and scaling. Distribution-based descriptors, therefore more robust. Human Motion from Mono-Camera (cont.) Generative Methods: Human Motion from Mono-Camera (cont.) Dimension Reduction: Human Motion from Mono-Camera (cont.) . GP-based dimensionality reduction. . Latent manifold controls both observation and state. . Associated dynamics in latent space. . Small data set for special motion type. . All parameters are learned by maximizing likelihood of the training data. . Allow fast inference. Dimension Reduction: - Shared Latent Dynamical Model Human Motion from Mono-Camera (cont.) Sta