A novel algorithm for designing optimized orthonormal transform-matrix codebooks for adaptive transform coding of a non-stationary vector process is proposed. This algorithm relies on a block-wise stationary model of a non-stationary process and finds a codebook of transform-matrices by minimizing the end-to-end mean square error of transform coding averaged over the distribution of stationary blocks of vectors. The algorithm, which belongs to the class of block-coordinate descent algorithms, solves an intermediate minimization problem involving matrix-orthonormality constraints in a computationally efficient manner by mapping the problem from the Euclidean space to the Stiefel manifold. As such, the algorithm can be broadly applied to any adaptive transform coding problem. Preliminary results obtained with inter-prediction residuals in an H265 video codec are presented to demonstrate the advantage of optimized adaptive transform codes over non-adaptive codes based on the standard DCT.