**5. Experimental Results**

Two datasets are used to validate the proposed algorithm. The first dataset is the Sports and Daily Activities (SDA) dataset [10], in which eight subjects were enrolled to perform 19 daily and sports activities while wearing five Xsens MTx (Details of Xsens MTx can be found in http://www. xsens.com/en/general/mtx)) IMUs on their torso, left arm, right arm, left thigh, and right thigh, all the sensors measure the acceleration and angular rate of the body parts where the sensors are placed. The five sensor placements are determined in this dataset because the involved activities are not only lower limb locomotion activities but also include static activities and upper limb activities, such as sitting, lying, rowing, and playing basketball... So they used the 5 sensors to collect the motion data from different parts of the body. However, in this paper, the proposed algorithm is designed for recognizing the lower limb locomotion activity with periodic gait or leg cycle performed by healthy people. Since we only care about the movement of the lower limbs and healthy people have a symmetric motion of the two legs, then it is possible to use only one sensor placed on either the left or right leg to recognize the considered activities. Therefore, we only use the sensor placed on the right thigh to validate our algorithm. The sensor sampling rate was set to 25 Hz, the acceleration sensing range was set to ±18 g, the angular rate sensing range was set to ±1200◦/s. Because the objective of the proposed algorithm is to detect lower limb locomotion activities that have gait cycle or leg cycle, while the 19 activities consists of both lower limb locomotion activities with and without the cycles, then only 11 suitable activities out of the total are selected in this work: walk in parking lot, walk on treadmill with incline, walk on treadmill on flat, stair descent, stair ascent, run on treadmill, jump, exercise on stepper, exercise of cycling in vertical position, exercise of cycling in horizontal

position, exercise on cross trainer. These 11 locomotion activities of SDA dataset are referred to as D1A1 to D1A11 in the remaining of this paper. In the dataset, the subjects performed each activity for about 5 min separately, and each activity was divided into 60 segments of 5 s. Therefore, there are 60 × 8 = 480 segments for each activity. In order to make the dataset available for our algorithm, we firstly combined the 60 segments of one activity from one subject to recover the 5 min activity, then combined the data of the same activity from all the subjects to form 40 min for each activity, the final data was obtained by combining the 11 activities. Thus, the duration of the data is 11 × 40 = 440 min.

There are only 7500 samplings for each experiment of SDA, and the data length is not long enough to use on-line EM recognition. Therefore, we utilize the second dataset for the validation of the proposed on-line EM algorithm. This second dataset, described in [7], is called Locomotion of Foot-mounted IMU (LMFIMU) dataset (the dataset and its details are available on the website: https://github.com/unilee/TMC\_LowerLimbActs). Ten subjects were enrolled to perform a specific experiment that lasts nearly 30 min with a Shimmer3 (Shimmer3 GSR+, details at the manufacturer's site http://www.shimmersensing.com/images/uploads/docs/ConsensysPRO\_Spec\_Sheet\_v1.1.0. pdf) IMU mounted on the right shoe. The sensor is placed on the shoe in LMFIMU but not the thigh as in SDA dataset is because both feet and thighs have a periodic pattern when people are performing the lower limb locomotion activities. In fact, the sensor can be placed at anywhere of the lower limbs in this study. So foot is chosen in the LMFIMU dataset to show that the proposed algorithm is not restricted to only one sensor position. Because of the enrolled 10 healthy subjects, left foot and right foot have the same behavior, then the sensor is placed on the right shoe of each subject. Each subject was asked to perform one experiment, which contained two identical sections of a sequence of four locomotion activities: walking, running, stair ascent and stair descent. Therefore, the performance of the second section will be improved compared to the first section, if the on-line algorithm can gradually learn the activity pattern of the subject. The four locomotion activities are referred to as D2A1 to D2A4 in the rest of this paper. The sensor sampling rate was set to 100 Hz, so the data length is long enough for the on-line EM algorithm. The sensing range of the acceleration and angular rate are ±8 g and ±1000◦/s, respectively.

The proposed SemiTMC-GMM model is compared with TMC-GMM to see the advancement of semi-Markov structure in recognizing lower limb locomotion activities. GMM is implemented by different *κ* to see the impact of the GMM components number that has on recognition accuracy.
