The paper introduces the Condensation algorithm, a stochastic method for tracking curves in dense visual clutter. Unlike Kalman filters, which are limited to Gaussian densities and cannot represent multiple hypotheses, Condensation uses factored sampling to propagate a full probability distribution over time, enabling robust tracking of agile motion in clutter. The algorithm combines learned dynamical models with probabilistic inference to represent the distribution of possible object positions and shapes. It operates in near real-time and outperforms Kalman filtering in handling multi-modal conditional densities, which are common in cluttered environments. The problem of tracking curves in clutter involves modeling the shape and motion of foreground objects against a complex background. Effective methods for shape and motion modeling have been developed in computer vision, but traditional approaches struggle with multi-modal distributions. The Condensation algorithm addresses this by using a stochastic approach that samples from the probability distribution of possible object states, allowing it to represent complex, non-Gaussian distributions. Sampling methods are used to estimate the posterior distribution of object states given observations. The algorithm uses factored sampling, where a set of samples is generated from the prior distribution and then weighted by the likelihood of the observations. This allows for efficient estimation of properties like mean and variance of the posterior distribution. Kalman filters are not suitable for tracking continuous curves in clutter because they assume Gaussian distributions and cannot handle multi-modal data. The Condensation algorithm, on the other hand, uses a stochastic differential equation model for object motion and a factored sampling approach to propagate the conditional density over time. This enables the algorithm to track complex, non-rigid motions in cluttered environments. The algorithm is applied to track curves in video streams, using probabilistic models for both the dynamics of the object and the measurement process. The model parameters are learned from training sequences, and the algorithm is initialized with a set of samples. The algorithm is tested on various scenarios, including tracking multi-modal distributions, rapid motions through clutter, and complex jointed objects. It is shown to outperform Kalman filters in these tasks, particularly in cluttered environments where multi-modal distributions are common. The Condensation algorithm is a powerful method for tracking curves in clutter, combining statistical factored sampling with a stochastic differential equation model. It is able to represent complex, non-Gaussian distributions and is effective in tracking agile motion in cluttered environments. The algorithm is demonstrated to be robust and efficient, with performance improving as the number of samples increases. It is able to track complex, non-rigid motions in cluttered environments, making it a valuable tool for visual tracking applications.The paper introduces the Condensation algorithm, a stochastic method for tracking curves in dense visual clutter. Unlike Kalman filters, which are limited to Gaussian densities and cannot represent multiple hypotheses, Condensation uses factored sampling to propagate a full probability distribution over time, enabling robust tracking of agile motion in clutter. The algorithm combines learned dynamical models with probabilistic inference to represent the distribution of possible object positions and shapes. It operates in near real-time and outperforms Kalman filtering in handling multi-modal conditional densities, which are common in cluttered environments. The problem of tracking curves in clutter involves modeling the shape and motion of foreground objects against a complex background. Effective methods for shape and motion modeling have been developed in computer vision, but traditional approaches struggle with multi-modal distributions. The Condensation algorithm addresses this by using a stochastic approach that samples from the probability distribution of possible object states, allowing it to represent complex, non-Gaussian distributions. Sampling methods are used to estimate the posterior distribution of object states given observations. The algorithm uses factored sampling, where a set of samples is generated from the prior distribution and then weighted by the likelihood of the observations. This allows for efficient estimation of properties like mean and variance of the posterior distribution. Kalman filters are not suitable for tracking continuous curves in clutter because they assume Gaussian distributions and cannot handle multi-modal data. The Condensation algorithm, on the other hand, uses a stochastic differential equation model for object motion and a factored sampling approach to propagate the conditional density over time. This enables the algorithm to track complex, non-rigid motions in cluttered environments. The algorithm is applied to track curves in video streams, using probabilistic models for both the dynamics of the object and the measurement process. The model parameters are learned from training sequences, and the algorithm is initialized with a set of samples. The algorithm is tested on various scenarios, including tracking multi-modal distributions, rapid motions through clutter, and complex jointed objects. It is shown to outperform Kalman filters in these tasks, particularly in cluttered environments where multi-modal distributions are common. The Condensation algorithm is a powerful method for tracking curves in clutter, combining statistical factored sampling with a stochastic differential equation model. It is able to represent complex, non-Gaussian distributions and is effective in tracking agile motion in cluttered environments. The algorithm is demonstrated to be robust and efficient, with performance improving as the number of samples increases. It is able to track complex, non-rigid motions in cluttered environments, making it a valuable tool for visual tracking applications.