This paper proposes a new method for estimating 3D camera orientation from video sequences in urban environments under the Manhattan world (MW) assumption, which posits that many edges are aligned with three orthogonal directions. The method avoids feature detection and correspondence, instead exploiting the MW assumption to reduce the search space for orientation estimation. Key contributions include defining equivalence classes of equiprojective orientations, introducing a small rotation model to capture smooth camera motion, and decoupling elevation and twist angle estimation from the compass angle. A probabilistic sequential estimation method is developed based on an MW likelihood model, significantly reducing the search space for each orientation estimate. The method is demonstrated using real video sequences, showing effective performance even in low-quality scenarios with many spurious edges. The approach avoids standard intermediate steps like feature detection and correspondence, and experimental results show its ability to handle low-quality video sequences. The method uses a probabilistic framework with a maximum a posteriori (MAP) criterion, incorporating the MW and small rotation (SR) assumptions to enable efficient and accurate orientation estimation. The algorithm is implemented in MATLAB and achieves real-time performance, with plans for a C implementation to improve frame rate. The method is validated on outdoor MPEG-4 video sequences, demonstrating its effectiveness in estimating camera orientation despite challenging conditions.This paper proposes a new method for estimating 3D camera orientation from video sequences in urban environments under the Manhattan world (MW) assumption, which posits that many edges are aligned with three orthogonal directions. The method avoids feature detection and correspondence, instead exploiting the MW assumption to reduce the search space for orientation estimation. Key contributions include defining equivalence classes of equiprojective orientations, introducing a small rotation model to capture smooth camera motion, and decoupling elevation and twist angle estimation from the compass angle. A probabilistic sequential estimation method is developed based on an MW likelihood model, significantly reducing the search space for each orientation estimate. The method is demonstrated using real video sequences, showing effective performance even in low-quality scenarios with many spurious edges. The approach avoids standard intermediate steps like feature detection and correspondence, and experimental results show its ability to handle low-quality video sequences. The method uses a probabilistic framework with a maximum a posteriori (MAP) criterion, incorporating the MW and small rotation (SR) assumptions to enable efficient and accurate orientation estimation. The algorithm is implemented in MATLAB and achieves real-time performance, with plans for a C implementation to improve frame rate. The method is validated on outdoor MPEG-4 video sequences, demonstrating its effectiveness in estimating camera orientation despite challenging conditions.