We present and evaluate a microcontroller-optimized limited-memory implementation of a Warping Longest Common Subsequence algorithm (WarpingLCSS). It permits to spot patterns within noisy sensor data in real-time in resource constrained sensor nodes. It allows variability in the sensed system dynamics through warping; it uses only integer operations; it can be applied to various sensor modalities; and it is suitable for embedded training to recognize new patterns. We illustrate the method on 3 applications from wearable sensing and activity recognition using 3 sensor modalities: spotting the QRS complex in ECG, recognizing gestures in everyday life, and analyzing beach volleyball. We implemented the system on a low-power 8-bit AVR wireless node and a 32-bit ARM Cortex M4 microcontroller. Up to 67 or 140 10-second gestures can be recognized simultaneously in real-time from a 10Hz motion sensor on the AVR and M4 using 8mW and 10mW respectively. A single gesture spotter uses as few as 135
W on the AVR. The method allows low data rate distributed in-network recognition and we show a 100 fold data rate reduction in a complex activity recognition scenario. The versatility and low complexity of the method makes it well suited as a generic pattern recognition method and could be implemented as part of sensor front-ends.