The idea that DTW is a fix for variable length is a common misconception. You need "uniform scaling" instead. Google the paper below, it explains it, and links to free code (and tests on one or two different accelerometer datasets, with very good results) Bing Hu, Yanping Chen and Eamonn Keogh. Time Series Classification under More Realistic Assumptions. SDM 201
What's next? You probably see all this, and then you get frustrated that “data science” is hard and that the only “solution” is “machine learning”. That's OK. So let's start with a quick reminder that there are many approaches to time series analysis. But there's no one-size-fits-all approach. We need a good understanding choosing and use it. We need to select the right algorithm for the job. Furthermore, we are still using the “time series analysis” paradigm.