The recognition of a finger movement has been a lot studied as the method of text input or UI manipulation for wrist-worn devices. In most cases, there has been some problem when users should wear a ring or a band to recognize their hand gestures. In this paper, we implement an efficient method for recognizing hand movements without additional devices on the fingers. For this object, finger movement information can be yielded by a video camera attached to the underside of the wrist. To discriminate fingers from arbitrary backgrounds in the wide range of illumination, it can be used that the distance of backgrounds is different from that of fingers and the amount of illumination is quite different according to the distance from the light source to the objects. For gesture recognition, input data into neural network can be acquired by dividing an image into 10$\times$10 tiles and averaging all the pixels in the tiles. By using subsampled images, a feedforward neural network is trained, and it is used as a classifier for consecutive image frame. Neural network``s output is mapped into a message for UI, which means neural network determines a finger. MultiInput method uses a total reflected light for detecting a finger``s degree of bending. This method in combination with multimodal feedback enables rhythm based text entry. To verify the effectiveness and efficiency of FreeFingers as a pointing device and a text entry device, a series of experiments were conducted and the results were compared with reference device.