2.3 Conclusions and Future Work
3.2.5 Evaluation of Sensor positions
The position on the body on which the accelerometer or gyroscope is worn can create a significant effect on the signal generated. While each part of the body undergoes periodic motion as steps are taken, the specific motion signals generated can vary significantly. These variations can affect step detection accuracy, as shown in Table 3.13. Because each RCA value is greater than 1, it can be seen that, on average, the algorithms and gaits recorded in this dataset tend to cause an overestimation of steps regardless of where the sensor is worn. However, with a RCA of 2.03 for the wrist worn sensor and an RCA of 1.46 for the ankle worn sensor, it is clear that the step overestimation problem can be significantly decreased if the sensor is worn on the ankle rather than the wrist. The hip worn accuracy is 1.56, which is more similar to the ankle than the wrist. This indicates that the wrist shows the most dramatic increase in step overestimation, which is likely in part due to the motion generated in non-stepping activities. It is far more common for the wrists to be in motion than for the ankle or hip to be in motion when not taking steps.
3.3
Conclusions and Future Work
This chapter described the analysis and categorization of pedometer algorithms, the meth- ods used for reimplementing three prior algorithms, and the analysis of the three algorithms with regard to parameter training, sensor position, gait type, and evaluation metric. Previous pedometer
algorithms can be categorized based on the base mechanism used to detect steps. This categoriza- tion process revealed that the majority of published step detection algorithms rely on peak or zero crossing (based on an adaptive threshold) detection. Based on this finding, two recent (2010 or more recent) algorithms, a peak detection and a threshold based zero crossing detection algorithm, were reimplemented and tested on the collected dataset. In addition, an autocorrelation based algorithm was implemented for comparison. Algorithm evaluation was performed across 5 axes; algorithm, parameter training, sensor position, gait type, and evaluation metric.
Results were obtained for each axis by collapsing the 108 accuracy results calculated for each combination of the 5 axes examined. Parameter space for each algorithm was searched to find the combination of parameters yielding the highest running count accuracy for the regular gait activity. These trained parameters outperformed the default parameters from the original publications in 40 of 54 combinations of algorithm, gait type, sensor position, and evaluation metric. In order to simplify further evaluation, the trained parameters were used in further analysis. Overall, no single algorithm produced the highest accuracy across all combinations of sensor position and gait type. Thus, ideal algorithm for use in a specific situation is dependent on the combination of gait type and sensor position used. With regard to gait type, as gait type moved from regular to unstructured, accuracy generally decreased regardless of algorithm or sensor position. In addition, accuracy tended to be highest when sensors were worn on the ankle, slightly less accurate when worn on the waist, and significantly less accurate when worn on the wrist. The two evaluation metrics examined were running count accuracy (RCA) and step detection accuracy (SDA), as represented by an F1 score. In general, the same trends across gait type and sensor position were seen, regardless of which metric was used. However, when evaluating differences in algorithm accuracy, the results from the two metrics differed, with a better RCA produced by the autocorrelation algorithm and a better SDA produced by the peak detection algorithm.
A trend of overestimating steps was seen in almost all cases for the three pedometer algo- rithms implemented in this work. This trend is particularly interesting when compared to the error seen in the Fitbit step detection from the previous chapter. The overall step count recorded by the Fitbit underestimated steps regardless of gait. This contrast between implemented algorithm overestimation and Fitbit underestimation of steps indicates that the algorithm used by Fitbit is either better tuned to filter out non-step activities (at the cost of detecting some steps) or uses a different core algorithm.
Future work includes searching parameter space for each algorithm to find optimal parame- ters for each combination of sensor position, gait, and evaluation metric. While the results indicate that training parameters to the specific dataset does significantly improve accuracy, it would be interesting to determine if training for a specific combination of sensor position, gait, and metric could identify a set of parameters which works best overall. In addition, future work could examine accuracy for individual participants to identify if specific gait patterns or sensor orientations cause significant changes in accuracy. It would also be interesting to implement a wider variety of pedome- ter algorithm categories, including gyroscope based algorithms or algorithms based in the frequency domain.
Chapter 4
Improving Pedometer Performance
Through Detection of Gait Type
Pedometer algorithm performance can vary significantly based on the location of the body that the sensor is worn and type of gait being performed by the wearer. If the sensor position and gait type were known, parameters and algorithms could be selected in order to maximize accuracy. Prior research has focused on identifying the position on which the sensor is worn, but no works have attempted to detect the more interrupted gaits that are frequent in daily life. We have designed experiments to elicit gaits common when exercising, walking around a building, and moving around a room, and we developed a classifier to identify which gait is being performed and select the most accurate algorithm accordingly.
4.1
Methods
This section describes the methods used to identify which gait is being performed and to select and switch to the most accurate pedometer algorithm based on the gait detected. Because no prior works have attempted to detect the types of gaits present when walking around a building or around a room, novel techniques were developed, and future of analyses can build on the concepts used in this work. When comparing gaits, it became clear that one feature which consistently differs between regular, semi-regular, and unstructured gait is the regularity of motion.
Regular gait, representing exercise activities like walking around a track or on a treadmill, can be defined by a consistent pace followed for several minutes at a time. In semi-regular gait, representing the motion patterns seen when walking around a building, there are some periods of regular motion (while walking down a hallway, for example) which are broken up when opening doors and can include changes in the rate of motion when walking up or down stairs. In addition, walking around a building includes some periods of walking around a room, which may also elicit a different pace. Thus, while there is some regularity to the motion overall, the pace is spread out over a wider range and is more irregular. In unstructured gait, representative of motions seen when walking around a room (for example, cooking dinner, hunting for keys, or vacuuming), there are periods of no motion, periods of motion without steps, and some periods of walking during which a regular stride is rarely reached. The differences in the regularity of the motion, and therefore acceleration signals, indicates that gait can be identified using the frequency domain.