This paper presents an automated routine for calibrating a fingernail imaging system, with the intent of predicting fingerpad forces. The system uses a magnetic levitation haptic device to apply forces to the human fingerpad while recording images of the nail and surrounding skin. A novel force controller is implemented to interact stably with the human fingerpad. The data are used to calibrate a principal component regression model relating pixel intensity to 3-D force. Using data from this automated routine, this model simultaneously predicts 3-D force with an RMS error of 0.56 ± 0.03 N (7.7% of the full range of forces).

This paper investigates the effect of two different parameters on the registration and force prediction accuracy of using Active Appearance Models (AAM) to align fingernail images. First, the color channel used to form the AAM is varied between (1) an averaged grayscale image, (2) the red channel, (3) the green channel and (4) the blue channel. Second, the number of landmark points used to create the AAM is varied between 6 and 75. The color channel is found to have an effect on the registration accuracy and the force prediction error. The green, blue and grayscale images are approximately equivalent, while the red images have a larger error across all metrics used. The number of landmark points may be reduced to 25 with no significant effect on either the registration accuracy or the force prediction error, though further reduction has shown some effect. With this information, a simpler registration model can be used that requires fewer calculations.

This paper demonstrates the effectiveness of registering fingernail images using Active Appearance Models (AAM). A shape model is formed using two contours, one to trace the edge of the fingernail and a second to outline the edge of the finger. This model is used to create an AAM, which is then used to register several hundred images. Using the images and associated force measurements, three different models relating pixel intensity to finger contact force are formed. This process is repeated for eight subjects. These image-force models could then be used to measure the finger contact force without obstructing the haptic sense or instrumenting the object being touched, allowing for more natural interaction between human and robot. Two of the models are found to predict force with an RMS error of 0.3N normal force (4% of the full range) and 0.23N shear force (3% of the full range). These two models are also shown to be nearly independent of image resolution, with no significant difference in the results for resolutions between 10-by-10 and 50-by-50. The simpler of the two models relies on the Texture Parameters generated by the AAM. Since these parameters are already calculated as part of the registration process, this removes an additional calculation step and could lead to improved real-time processing.