Laser imaging radar can obtain point cloud data reflecting the three-dimensional position of the target, directly estimate the three-dimensional attitude angle of the target, and is an important parameter for feature extraction and target registration. To realize the three-dimensional attitude estimation of scenes, an optimized three-dimensional attitude estimation algorithm(OPDVA) based on point normal vector (PDVA) was proposed to solve the problem of large deviation of the positive vector representing the coordinate axis of scene coordinate system(SCS) in real scenes. In this method, remove point normal vectors in other directions in the cluster by RANdom SAmple Consensus (RANSAC) plane model was removed, and the corresponding normal vectors of the optimal fitting plane were the revised SCS coordinate axes. Using rotational transformation and resampling techniques, 3 groups of real scene range image were experimented with rectangular bounding box method, PDVA and OPDVA respectively. The experimental results show that the OPDVA method is superior to the other two methods in pose estimation. The error of pose estimation does not exceed 4°, and it is also suitable for occlusion scenarios.