This paper presents a probabilistic approach to image segmentation in video sequences, focusing on the task of identifying moving objects, such as vehicles, in traffic surveillance. The traditional "background subtraction" method, which subtracts a time-averaged background image from the current image to detect moving objects, is criticized for its inability to distinguish between moving objects and shadows, and for failing with slow-moving objects. The proposed method uses a probabilistic model to classify each pixel as part of a moving object, shadow, or background. This model is learned using an unsupervised, incremental version of the EM algorithm, which allows for real-time processing and automatically updates the model based on the likelihood of pixel membership. The key idea is to model the appearance of each pixel as a mixture of Gaussians, where each Gaussian represents a different class (road, shadow, or vehicle). This approach allows for more accurate classification of pixels, especially in the presence of shadows and slow-moving objects. The method is applied in the Roadwatch traffic surveillance project, where it is expected to improve vehicle identification and tracking. The paper describes the background subtraction method, which involves computing a long-term average background image and identifying moving objects by comparing the current image to the background. However, this method is limited by changes in lighting conditions and the inability to distinguish between shadows and moving objects. The proposed method addresses these issues by using a probabilistic model that can adapt to changes in the scene and accurately classify pixels. The paper also discusses the use of the EM algorithm for learning pixel models, with an incremental version that allows for real-time processing. This approach is more effective than traditional methods in handling shadows and slow-moving objects. The results show that the proposed method significantly improves the detection and tracking of vehicles in traffic surveillance. The paper concludes that a probabilistic approach, combined with an unsupervised learning algorithm, can significantly improve the detection of moving objects in video sequences. Future work includes incorporating more background knowledge into the models to improve performance and handle more complex scenarios.This paper presents a probabilistic approach to image segmentation in video sequences, focusing on the task of identifying moving objects, such as vehicles, in traffic surveillance. The traditional "background subtraction" method, which subtracts a time-averaged background image from the current image to detect moving objects, is criticized for its inability to distinguish between moving objects and shadows, and for failing with slow-moving objects. The proposed method uses a probabilistic model to classify each pixel as part of a moving object, shadow, or background. This model is learned using an unsupervised, incremental version of the EM algorithm, which allows for real-time processing and automatically updates the model based on the likelihood of pixel membership. The key idea is to model the appearance of each pixel as a mixture of Gaussians, where each Gaussian represents a different class (road, shadow, or vehicle). This approach allows for more accurate classification of pixels, especially in the presence of shadows and slow-moving objects. The method is applied in the Roadwatch traffic surveillance project, where it is expected to improve vehicle identification and tracking. The paper describes the background subtraction method, which involves computing a long-term average background image and identifying moving objects by comparing the current image to the background. However, this method is limited by changes in lighting conditions and the inability to distinguish between shadows and moving objects. The proposed method addresses these issues by using a probabilistic model that can adapt to changes in the scene and accurately classify pixels. The paper also discusses the use of the EM algorithm for learning pixel models, with an incremental version that allows for real-time processing. This approach is more effective than traditional methods in handling shadows and slow-moving objects. The results show that the proposed method significantly improves the detection and tracking of vehicles in traffic surveillance. The paper concludes that a probabilistic approach, combined with an unsupervised learning algorithm, can significantly improve the detection of moving objects in video sequences. Future work includes incorporating more background knowledge into the models to improve performance and handle more complex scenarios.