Improving the measurement accuracy of distance and positioning for capacitive proximity detection in human-robot interaction

Capacitive proximity sensors have been widely used in human-robot interaction, but suffer from the problem of the imperfect measurement accuracy in the nonlinear detection range. Meanwhile, sensor calibration makes the measurement tedious, time consuming, and is difficult to be used in 3D positioning. In this paper, we use a hand as the sensing target and propose a new distance estimation and positioning method, which combines with capacitive sensing configurations of a sensor array and machine learning. Fourteen sensing configurations are designed to generate 14 capacitive sensors to detect the target, and the capacitances are used as the input vector of the machine learning models. We compare the four machine learning algorithms (i.e. fitting method, support vector regression (SVR), radial basis function (RBF) network and BP neural network (BPNN)) for hand distance estimation. The experimental results show the RBF offers the best performance that the root mean squared error (RMSE), mean absolute error (MAE) and mean relative error (MRE) are 0.59 cm, 0.45 cm and 5.32%. Due to the lack of a proper 3D positioning device to verify the target positions, all tests for positioning are grounded on the simulation data using COMSOL, which can mimic the real world applications with high accuracy. The testing results are compared regarding the use of a Y-shaped sensor arrangement, and show that our proposed method performs better than the method with a Y-shaped sensor arrangement, the accuracy is improved by more than 98.8% and the BPNN model achieves the highest accuracy.