This paper presents a probabilistic method for tracking 3D articulated human figures in monocular image sequences. Within a Bayesian framework, a generative model of image appearance is defined, along with a robust likelihood function based on image graylevel differences and a prior probability distribution over pose and joint angles that models human motion. The posterior probability distribution over model parameters is represented using a discrete set of samples and is propagated over time using particle filtering. The approach extends previous work on parameterized optical flow estimation by exploiting a complex 3D articulated motion model. It also extends previous work on human motion tracking by including a perspective camera model, modeling limb self-occlusion, and recovering 3D motion from a monocular sequence. The explicit posterior probability distribution represents ambiguities due to image matching, model singularities, and perspective projection. The method relies only on a frame-to-frame assumption of brightness constancy and is able to track people under changing viewpoints, in grayscale image sequences, and with complex unknown backgrounds. The paper presents a Bayesian approach to tracking 3D articulated human figures in monocular video sequences. The human body is represented by articulated cylinders viewed under perspective projection. A generative model is defined in terms of the shape, appearance, and motion of the body, and a model of noise in the pixel intensities. This leads to a likelihood function that specifies the probability of observing an image given the model parameters. A prior probability distribution over model parameters depends on the temporal dynamics of the body and the history of body shapes and motions. With this likelihood function and temporal prior, the posterior distribution over model parameters at each time instant, given the observation history, is formulated. The estimation of 3D human motion from a monocular sequence of 2D images is challenging due to non-linear dynamics of the limbs, ambiguities in the mapping from 2D image to 3D model, similarity of appearance of different limbs, self-occlusions, kinematic singularities, and image noise. These difficulties result in a multi-modal posterior probability distribution over model parameters. The paper also discusses related work on human motion estimation, including image cues, body representations, temporal models, and estimation techniques. The approach here focuses on the estimation of 3D articulated motion from 2D image changes, exploiting recent work on the probabilistic estimation of optical flow using particle filtering.This paper presents a probabilistic method for tracking 3D articulated human figures in monocular image sequences. Within a Bayesian framework, a generative model of image appearance is defined, along with a robust likelihood function based on image graylevel differences and a prior probability distribution over pose and joint angles that models human motion. The posterior probability distribution over model parameters is represented using a discrete set of samples and is propagated over time using particle filtering. The approach extends previous work on parameterized optical flow estimation by exploiting a complex 3D articulated motion model. It also extends previous work on human motion tracking by including a perspective camera model, modeling limb self-occlusion, and recovering 3D motion from a monocular sequence. The explicit posterior probability distribution represents ambiguities due to image matching, model singularities, and perspective projection. The method relies only on a frame-to-frame assumption of brightness constancy and is able to track people under changing viewpoints, in grayscale image sequences, and with complex unknown backgrounds. The paper presents a Bayesian approach to tracking 3D articulated human figures in monocular video sequences. The human body is represented by articulated cylinders viewed under perspective projection. A generative model is defined in terms of the shape, appearance, and motion of the body, and a model of noise in the pixel intensities. This leads to a likelihood function that specifies the probability of observing an image given the model parameters. A prior probability distribution over model parameters depends on the temporal dynamics of the body and the history of body shapes and motions. With this likelihood function and temporal prior, the posterior distribution over model parameters at each time instant, given the observation history, is formulated. The estimation of 3D human motion from a monocular sequence of 2D images is challenging due to non-linear dynamics of the limbs, ambiguities in the mapping from 2D image to 3D model, similarity of appearance of different limbs, self-occlusions, kinematic singularities, and image noise. These difficulties result in a multi-modal posterior probability distribution over model parameters. The paper also discusses related work on human motion estimation, including image cues, body representations, temporal models, and estimation techniques. The approach here focuses on the estimation of 3D articulated motion from 2D image changes, exploiting recent work on the probabilistic estimation of optical flow using particle filtering.