Warm-up Actions Detection Based On Sensors Built-in Smartphone

Running is a good way to keep healthy and relax,while many runners suffer from injuries because of ignoring the importance of warm-up.Inspired by the phenomenon that more and more people run with smartphones tied to their arms,we propose a novel system named iRun to alert people to warm up before running.iRun is based on the sensors built in most off-the-shelf smartphones like accelerom-eter,and it's core module is about the accurate identification of warm-up actions.In order to achieve this goal,we first determine the warm-up-action set,then recruite vol-unteers to collect data and select features that can represent the characteristics of warm-up actions.By carefully designing the features vector which contains features from multi-domains and doing a series of experiments to decide the slide window size and classifier,we can achieve 92.2%true positive(TP)rate in average to distinguish every warm-up action from other movements like running,walking,going upstairs,etc.Although this result is excellent,it is hard to get further improvement because there are no clear hints on valid features.Inspired by recent outstanding works on neural networks,we designed data processing flows based on recurrent neural network(RNN)and convolutional neural network(CNN),both of which can make the test accuracy higher than 95.0%with fine-tuned parameters.The main contributions of this paper are as follows:(1)We collect and determine a set of warm-up-action that are common and effec-tive in daily exercise,and those actions are familiar to runners,especially the amateurs.Unlike other systems that use special wearable devices,which are difficult to be widely equipped by runners in short time,iRun is based on the built-in sensors of most off-the-shelf smartphones,making it possible to be immediately put into commission;(2)To distinguish the warm-up actions,we extracted a set of discriminative fea-tures in different domains,most of which require relatively few computing resource.This ensures that we can get accurate,real-time results with RandomForest;(3)To obtain more accurate recognition results,we designed the data processing flows based on based on recurrent neural network(RNN)and convolutional neural net-work(CNN).The data processing flow based on RNN is robust and has fewer parame-ters,the one based on CNN can produce more accurate results.