Sensor-based gait analysis is a valuable tool in diagnosis and assessment of Parkinson's disease. Especially for large data sets, efficient analysis pipelines are required. Pre-segmentation of long time series into chunks of interest is a possible approach to increase efficiency. Therefore, we developed an unsupervised algorithm for the detection of gait sequences from continuous sensor signals. In the proposed method, gyroscope signals representing the angular rate of the feet are analyzed in the frequency domain using moving windows. A gait sequence was detected, if the frequency spectrum of a given window contained harmonic frequencies. The approach was tested on two data sets that differed in the ratio of clinical gait and cyclic movement tests. Sensitivity in both data sets was higher than 99% in a stride-to-stride comparison with ground truth. The specificity was measured with 76.1% (data set 1) and 94.5% (data set 2) for tests against sequences of other cyclic movements. In conclusion, the algorithm offers a reliable and efficient approach for the detection of gait sequences in time series data and is also promising for the application in long-term home-monitoring scenarios.