Human activity recognition (HAR) is encountered ina plethora of applications, such as pervasive health care systemsand smart homes. The majority of existing HAR techniquesemploys features extracted from symbolic or frequency-domainrepresentations of the associated data, whilst ignoring completelythe behavior of the underlying data generating dynamical system.To address this problem, this work proposes a novel self-tunedarchitecture for feature extraction and activity recognition bymodeling directly the inherent dynamics of wearable sensordata in higher-dimensional phase spaces, which encode staterecurrences for each individual activity. Experimental evaluationon real data of leisure activities demonstrates an improved recog-nition accuracy of our method when compared against a state-of-the-art motif-based approach using symbolic representations.