Indoor localization is a key enabler of smart environments, yet achieving high-accuracy positioning remains difficult due to severe multipath effects, signal distortion, and the limited ability of current deep models to generalize across heterogeneous wireless datasets. Most existing approaches overlook the intrinsic sparsity structure of wireless signal features, leading to unstable latent representations and reduced robustness in complex indoor scenarios. To bridge this gap, this paper proposes a non-smooth regularized autoencoder that embeds a sparsity-inducing non-smooth regularizer into the latent space, enabling the extraction of compact, noise-resilient, and interpretable spatial features from high-dimensional signal measurements. This design introduces a clear advancement over state-of-the-art autoencoder and regression-based models, which predominantly rely on smooth penalties and consequently struggle under strong non line of sight conditions. Evaluations on the UJIIndoorLoc benchmark show that the non-smooth regularized autoencoder achieves an average localization error of 6.78 m, outperforming conventional deep neural network baselines. Furthermore, experiments on a real-world CSI dataset collected in a 13.5 × 11 m² cluttered laboratory demonstrate a mean error of 1.73 m in a challenging non line of sight environment. These findings confirm that the proposed non-smooth regularized autoencoder effectively addresses core challenges in robust feature learning and offers strong generalization across diverse indoor settings, establishing it as a powerful framework for next-generation indoor localization systems.