6533b7dcfe1ef96bd1272920

RESEARCH PRODUCT

Dynamics-based action recognition for motor intention prediction

subject

description

Abstract Powered lower-limb prostheses presented in the previous chapter require a natural and easy-to-use interface for communicating amputee’s motor intention in order to select the appropriate motor program in a given context or simply to commute from an active (powered) to a passive mode of functioning. To be accepted by amputees, such an interface should (1) not put additional cognitive load on the end-user, (2) be reliable and (3) be minimally invasive. In this chapter we present one possible solution for achieving that goal: a robust method for autonomously detecting and recognizing motor intents from a wearable sensor network mounted on a sound leg. The sensor network provides a real-time stream of accelerometer, gyroscope and magnetometer data, while a machine learning algorithm detects human intentions and activates an appropriate control program. We will demonstrate how machine learning is to be used in this setting: from feature extraction and selection, to model training and validation. While we will adopt specific features and models throughout this chapter, the presented approach is general enough that other similar state-of-the-art techniques can be similarly adopted.