Contents:
Energy of 6 equally spaced frequency bands within the 64 bins of the FFT. The features extractor has been implemented using Octave v. The machine learning algorithm module acts as a classifier.
For this module, we have tried two popular algorithms J48 decision tree and Random Forest and found that the Random Forest algorithm [ 28 ] works better in this circumstance. Therefore, the Random Forest algorithm is used in our following experiments. The Random Forest algorithm creates several decision trees during training. In our experiments, the number of trees ranges from 40 to 95, and the number of nodes per tree goes from 15 to These numbers varies strongly with the number of features considered in the feature extractor.
The algorithm for training the Random Forest model is: Open image in new window. For classification, every new input from the testing set is run down all of the trees. The classification result is weighted average of all of the terminal nodes that are reached, providing the final decision: Open image in new window.
This work has been carried out using the Random Forest implementation included in the WEKA toolkit [ 8 ] weka configuration weka. This section describes the experiments conducted in this work. In first and second subsection, the main dataset and the evaluation methods used in this work are introduced. Third and fourth subsections shown the experiments carried out for the analysis of the type of sensor and the type of feature.
In fifth subsection, the final results on the main dataset are given. These recording include 13 inertial signals obtained from 9 on-body inertial measurement units located on different body parts. Each unit contains four sensors: an accelerometer, a gyroscope, a magnetometer and a quaternion sensor. The experiment consisted in performing a complete set of exercises: 33 physical activities, including warm up, fitness and cool down activities walking, jogging, cycling, jumping, etc.
This dataset also includes a Null-activity. This label has been assigned to other activities different from the 33 activities considered in this study , and also, the transitions between activities. Ideal - placement all sensors were placed by experts at their optimal place for classification. All the subjects recorded a session in these conditions 17 sessions. Self - placement every subject decides the positions of three sensors by themselves and the remaining sensors were situated by experts.
The number of three is considered a reasonable estimate of the proportion of sensors that may be misplaced during the normal wearing. Mutual - placement where several displacements were intentionally introduced by experts. Three out of the 17 volunteers were recorded for mutual-displacement scenario subjects 2, 5 and These three subjects recorded one session for every sensor configuration: for the case in which four, five, six or even seven out of the nine sensors are misplaced.
Considering the size, the number of activities and the different placement scenarios, we think that the REALDISP dataset is an appropriate dataset to evaluate the performance of the proposed HAR system. In this work, we apply two methods to evaluate the system performance. The first one is a tenfold random-partitioning cross-validation evaluation. For every experiment, one subset is used for testing and the other nine for training, considering a round-robin strategy. The final cross-validation result is the average along the 10 experiments.
This is the method used in the original paper [ 24 ]. However, our hypothesis is that this method suffers the problem that the data in both training and testing could contain information from a same subject, so the machine learning algorithm can learn not only physical activity characteristics but also some subject-dependent ones.
This aspect makes it hard to evaluate the system performance when facing a new subject. In order to verify this hypothesis, sect. Therefore, we propose the second method, a subject-wise cross-validation.
In this case, the same kind of cross-validation is done but on different subjects rather on automatically split parts: all data from the same user is considered for testing and the data from the remaining subjects for training. Since we have 17 subjects in the database, this experiment is repeated 17 times. The final experimental result is the average of accuracy and F-measure on all 17 sub-experiments weighted by the number of samples in every testing data.
The rest experiments are conducted with the subject-wise method. In this paper, for all the experiments, the confidence interval is lower than 0. In this section, we first conduct some experiments for HAR system configuration tuning in order to analyze how the evaluation, the sensor type, or the feature type influences the system performance. This result supports the hypothesis stated in the previous section: in the random-part evaluation, training and testing subsets could contain information from a same subject and this characteristic produces better classification results.
In the rest of the paper, we will only consider the subject-wise evaluation method more challenging situation. Same as the experiments on sensor types, here, we also consider the ideal-placement, removing the Null-activity.
For the sake of confidence, we repeat the experiments with different types of sensor. Experimental results on type of feature.
Gjoreski and Gams 11 subjects [ 56 ]. Moore, Jason W. Depending on the performance of the used equipment, the main difference is the sampling rate with which the sensor device is capable of perceive its surroundings. But, in this very endeavor, Peirce was striving to show in detail how the different branches of investigation can fruitfully draw upon, and appeal to, one another. Rosenberger, Robert, and Peter-Paul Verbeek, eds. Sekine M. Activity recognition on an accelerometer embedded mobile phone with varying positions and orientations.
This figure represents and compares the results for different kinds of features: temporal features and frequency features and using different types of sensors. As a conclusion, it is clear that using the signals from magnetometer and the time-based features is currently the best system configuration.
By including all sensors, we obtain even higher system accuracy: Mean removal: Subtract the mean value from each value in a feature or signal vector. Z-Score: Mean removal first, and divide each value by its standard deviation. Histogram equalization: Consider all the values in a gray-scale, and equalize its histogram. Vector normalization with mean normalization: Vector normalization followed by mean removal.
By applying the best experimental configuration described above, we conducted experiments using data from all signals and all sensors in all the three placement scenarios. Another aspect to comment is regarding the Null-activity. We truncated the Null-activity samples in order to make a fair comparison with the original paper. In this work, we have also done experiments including the Null-activity, which, in our opinion, is closer to a real situation. So, the problem now becomes more challenging: a class classification task.
Such low degradation is made possible due to the large number of features extracted and the suitable normalization method proposed in this paper.
Regarding the self-placement scenario, this table shows 8. The new feature extraction module shows a very good robustness against different sensor placements. For the mutual-placement scenario, the results are considerably low in both works baseline and this paper but the degradation obtained with the system proposed in this paper is considerably smaller compared to the baseline system: our system shows a better robustness.
This degradation is different depending on the number of mis-displacements 4, 5, 6, and 7. Final experimental result for self-placement and mutual-placement scenarios: using ideal-placement for training.
The results show that when being trained with ideal datasets and tested with mutual datasets, the system reaches a very good accuracy though the training and testing sets come from different placement scenarios. For example, for mutual4, the accuracy goes from These results support the hypothesis that the amount of data for training is an important factor in the system performance. In this case, the amount of data available in ideal-placement and self-placement scenarios is the same.
In the introduction, we commented two main applications of HAR: physical exercise monitoring and home care monitoring.
The recordings include daily morning activities: getting up from the bed, preparing and having breakfast a coffee and a salami sandwich and cleaning the kitchen latter. This dataset is a very popular HAR dataset on this research field. There is no constraining on the location or body posture in any of the scripted activities.
In the Drill run, subject must act in a predetermined activity sequence and, as for ADL1—ADL5, there is no restriction on the order and number of activities.
For each subject, there is information from three types of sensors: body-worn sensors, object sensors and ambient sensors. The on-body sensors include 7 multi-sensor inertial measurement units with another 12 3D acceleration sensors: signals in total. Since only body-worn sensors are concerned in the evaluation section of the original paper [ 31 ], the data from object and ambient sensors are truncated in the following experiments.