**3.5 Discussion**

Comparisons between the results of the embedded system and the reference show that the joint angles measured by the embedded system have limited differences compared to the reference system. The RMSE of the proposed system is between 2.4° and 4.0°. The signal shapes of both systems have great similarities. The CC of the signals between two systems is between 0.89 and 0.99.


#### **Table 1.**

*Comparison between the proposed system and the 3D-GA system.*

**Figure 6.** *Correlation (a) and Bland-Altman plot (b).*

From the results of the comparison (**Table 1**), **Figures 6** and **7**, the largest RMSE and the smallest CC are observed for the joint angles of the ankles. Three sources may induce these errors: percussion during walking, centripetal acceleration, and the cross-talk effect [28]. During the walk, the main percussions occur when the foot contacts the ground. Percussions influence the measurements of accelerometers and gyrometers; the filtering process used to reduce these influences does not fully eliminate them. Measurement errors can also be introduced by the cross-talk effect. During a straight walk, the ankles have more degrees of freedom than other joints. When positioning a sensor, there can be misalignments between the sensor and the segment of the human body. The consequence is that the movements in the planes other than the sagittal plane are combined in the measure. This problem can be solved by adding a calibration phase to align the sensor and human body [29]. This problem will be considered in a future study.

**149**

reliable results.

*An Embedded Gait Analysis System for CNS Injury Patients*

Standard for normal gait due to their high accuracy.

force sensor for the pathological gait, and this limits its use [30].

To conclude this section, despite the lack of a precise calibration process, the embedded system provides an estimate of the joint angles close to the references,

Data captured by the system are organized in time series; this does not allow to analyze or compare inter or between patients. In gait analysis, locomotion data are often partitioned with gait event and normalized in gait cycle to compare with each other. The size of data captured in everyday life condition is often huge. A method

Detection of gait events is a fundamental problem. All captured gait signals should be normalized in gait cycles so that they could be compared between each other. Calculations of spatiotemporal parameters are also based on the timing of gait events. So, the determination of the IC and FC event occurrences is compulsory.

the ground reaction force (GRF) or by processing of the signals from IMU.

In everyday life conditions, the gait event can be captured whether by measuring

GRF-based methods use foot switches (force sensitive resistor or mechanical switch) placed under toe and heel to determine IC and FC events. Gait event detection is quite easy with these methods. Due to the GRF, the foot switch state toggles both when the foot contacts the ground (Heel-Strike, IC event) and when the foot leaves the ground (Toe-Off, FC event). These systems are considered as the Gold

But in pathological gait, the IC event could not correspond to heel-strike and the FC event not often corresponds to toe-off as in normal gait. For example, with foot drop gait, the IC event may be the moment when the toe starts to contact the ground and FC event is the moment of toe-off. Therefore, it is difficult to place the

Several studies propose real-time or delayed real-time motion detection systems based on the use of acceleration and angular velocity signals of human body segments provided by IMU [30–36]. All the algorithms proposed in these studies consist of a set of rules that make it possible to identify several characteristics in the different gait signals. Currently, in everyday life gait event determination systems, the accelerometer seems to be the most used sensor [37]. The data it provides are often coupled with data from the gyrometer to get more

**Figure 7** illustrates the acceleration, sagittal angular velocity, and inclination signals of both feet on a healthy subject during normal walking. The signals observed on both feet have almost the same characteristics and shapes. The local extrema of acceleration and angular velocity signals are the most commonly used characteristics for determining gait events (IC and FC) [37]. The accuracy of methods presented in different studies may achieve between 11 and 165 ms [35, 36].

Between 86 and 98% of events are correctly detected for normal gait [37].

**Figure 8** shows the acceleration, angular velocity, and inclination signals of the two feet of a hemiparetic patient during a straight walk. Due to the asymmetry of the gait, the acceleration and angular velocity signals are significantly distinct between the healthy side and the paretic side. The amplitudes of the acceleration and angular velocity signals are much lower than those of a healthy person. Local maximum is strongly attenuated, especially the local maximum corresponding to the IC event in the angular velocity signal. In addition, the style of pathological walking varies greatly between stroke persons and there is a significant variability in the shape of the recordings. Given this specificity, it is difficult to use existing

which validates both the architecture and the data processing algorithm.

*DOI: http://dx.doi.org/10.5772/intechopen.83826*

to perform auto-partition is necessary.

**4. Determination of gait events**

**Figure 7.** *IMU signals and gait events of a healthy subject.*

To conclude this section, despite the lack of a precise calibration process, the embedded system provides an estimate of the joint angles close to the references, which validates both the architecture and the data processing algorithm.

Data captured by the system are organized in time series; this does not allow to analyze or compare inter or between patients. In gait analysis, locomotion data are often partitioned with gait event and normalized in gait cycle to compare with each other. The size of data captured in everyday life condition is often huge. A method to perform auto-partition is necessary.
