Skip to Main Content
While mobile phones are increasingly being used in activity recognition, tasks that require arm motion monitoring have so far not been studied on phone platforms. We leverage the fact that upper arm holsters are an increasingly popular way of wearing mobile devices during physical exercise to investigate the suitability of such platforms for arm dominated activity recognition. We focus on (1) user independent recognition from (2) a NULL class dominated, continuous data stream and (3) requiring no special care in device attachment (apart from being placed in a commercial holster). These are 3 key requirements for a realization in a real life mobile "App". We evaluate our methods on a gym exercises data set from 7 users that contains 11'000 individual repetitions of 10 different upper body exercises organized in 700 "sets" (=consecutive repetitions of the same exercise). On set level we achieve a user independent recognition of 93.6%. In over 90% of cases we can also count individual instances with an accuracy of ±20%.