In MIMO-OTFS systems, accurate CSI estimation with minimal pilot overhead is crucial. While guard-interval (GI)-based pilot patterns suppress data interference, they substantially increase overhead. To overcome this, we propose an interference-free, low-overhead channel estimation framework that exploits channel sparsity. Specifically, we introduce the delay-Doppler-angle basis expansion model (DDA-BEM) to reformulate the estimation task as a compressed sensing problem. Leveraging the sparsity in the DDA domain, we design a pilot pattern that inherently avoids data interference without requiring GIs. Simulation results demonstrate a 54% reduction in pilot overhead compared to state-of-the-art methods, while maintaining comparable estimation accuracy.