Hand gesture recognition is a promising research area often used in applications of human–computer interactions in the medical field. In this paper, we present a novel approach to differentiate gestures based on an arc-length parametrization and a curvature analysis of 3D trajectories. This new method called dynamic arc length warping (DALW) can outperform classic multi dimensional-dynamic time warping (MD-DTW) algorithm as it is invariant to sensor location and more tolerant to temporal distortions. Experimental validation of the algorithm is presented using different gestures and sensors in biomedical applications: an exoskeleton apparatus, surgical gestures captured by an instrumented laparoscopic device and finally, a birth simulator with an instrumented forceps. A basic perceptron multilayer neural network was implemented in order to perform the classification. Results involve an average increase of 7.14% in the classification rates by using DALW distance, compared to the classical MD-DTW.