A curved palm visuotactile sensor was proposed to address the challenges faced by existing robotic visuotactile sensors, as most of the sensors employ planar sensing structures and have the limitations of incomplete contact information acquisition and insufficient depth perception in complex 3D environments and delicate manipulation tasks. The innovative curved structural design that extended the effective sensing area into the 3D space significantly enhanced the tactile perception depth and contact information integrity. A point cloud reconstruction framework using the binocular stereo vision-based refractive ray tracing model was established to generate high-precision 3D tactile point cloud data. An efficient point cloud saliency detection algorithm that enabled intelligent recognition and real-time extraction of key contact regions was proposed to improve the efficiency and accuracy of data processing. Experimental results demonstrated that the structural design and the point cloud reconstruction framework of the proposed sensor exhibited strong robustness, achieving a 3D tactile point cloud perception accuracy of 0.20 mm, a point cloud reconstruction frame rate of 30 frames per second, and a maximum pressing depth of 10 mm. The sensor can precisely extract the salient contact point clouds in real time, providing robust technical support for robotic fine manipulation in complex environments.