Robots usually need to work in human environments and handle many different types of objects. There are two problems that make this challenging for robots: Human environments are typically cluttered and the multi-finger robotic hand needs to grasp and lift objects without knowing their weight, coefficient of static friction, and stiffness. Thus, this thesis combines vision and robot hand action to achieve reliable and accurate object grasping in a cluttered scene. An efficient algorithm for collision-free grasping pose generation according to a bounding box is proposed and the grasp pose will be further checked for its grasp quality. Finally, by fusing all available sensor data appropriately, an intelligent grasp system is achieved that is reliable and enough to handle various objects with unknown weight, friction, and stiffness.