Continuous Classification of Locomotor Transitions Performed under Altered Cutting Style, Complexity and Anticipation

Intent recognition strategies for lower-limb assistive technologies (e.g., prostheses and exoskeletons) are typically limited to transitions from one terrain to another (e.g., level ground to stairs) occurring under anticipated cognitive states and in a straight path. However, such systems might be unable to robustly handle unanticipated transitions and those that do not occur in a straight line, posing high fall and injury risks for individuals suffering from cognitive and/or physical impairments. Understanding how classification of human locomotion is influenced by changes of direction and user anticipation is important. In this study, we examined the performance of linear discriminant analysis in continuous classification of straight-line walking, 45°changes of direction (crossover and sidestep cuts) and cuts to stair-ascent (i.e., 'mixed' transitions), performed under varied anticipatory conditions of able-bodied individuals. Accelerographic and gyroscopic signals of the lower- and upper-body were used as signal inputs. Relatively high accuracy levels were observed during cut-to-stair, compared to cuts. In addition, classification of unanticipated maneuvers was compared for differing amounts of anticipated and unanticipated training data. It was observed that not more than two bouts of a given transition were needed to provide significant improvement in recognition of unanticipated mixed transitions. However, substantial (5 bouts) unanticipated information was needed to deliver improved performance of cuts (without stair). This study has implications for enhancing the design of recognition systems that adapt to changes of direction and user cognition. These outcomes could inform robust handling of 'unknown' circumstances and deliver an increased level of user safety and confidence.