Human Activity Recognition based on Mobile Phone Sensor

Abstract:

Recognition of human activity on the sensor is used in many aspects of everyday life. Face recognition, fingerprint sensor, GPS sensor, touch sensor, sensor tracker, and many other sensors are used. These sensors give us accurate results, and the results are comfortable for us so that the accelerometer performs various activities such as sitting, down, standing, limping, jogging, and upstairs.

INTRODUCTION:

This topic is entirely technology based. Various sensors are found in smartphones, such as an accelerometer, magnetometer, GPS, proximity sensor, ambient light sensor, camera, touch screen sensor, fingerprint sensor, barcode / QR code sensor, heart rate sensor, thermometer, A wide variety of device applications relies primarily on the detection of human behavior (HAR) in the workplace, such as patient and elderly monitoring, surveillance systems, robots learning, and military applications. The idea of an automatic HAR system depends on gathering measurements from some suitable sensors which are affected by selected attributes of human motion. Then some features are extracted depending on these measurements to be used in the process of training activity models. Which in turn will be used for later recognition of such activities.

Machine Learning & Methods:

Understand and make use of. Although machine learning is a field of computer science, it is distinct from traditional computing approaches. In traditional computation, algorithms are a set of directly designed instructions used by computers to calculate or solve problems. Computers can train data inputs rather than machine learning algorithms and use statistical analysis to generate values within a specific range.

Machine learning thus enables the development of models from experimental data for computers to simplify decision-making processes based on data inputs. Every technology consumer today has benefited from machine learning. Face-recognition technology allows social media platforms to help users tag and post friends’ images.

Optical character recognition software (OCR) converts text images to phone form. Machine learning-driven recommendation engines suggest what movies or TV shows to watch next based on user preferences. Machine learning-driven recommendation engines suggest what movies or TV shows to watch next based on user preferences. Consumers could soon have access to self-driving cars that rely on machine learning to navigate.

The most widely adopted machine learning approaches are supervised learning which trains algorithms based on examples of human-labeled input and output data, and unsupervised learning, which provides the algorithm with no labeled data to allow it to find structure within its input data.

Related Work

Locations and counts of sensors are very important issues that must be taken into account when designing a HAR system based on an accelerometer. Many settings were studied through the previous work, as seen in Table 1. Deferent body locations have been used from feet to chest with respect to the locations of wearable sensors. The selection’s relevance to the activities, however, plays an important role in specifying the sensor’s location. For example, Outpatient movements (such as driving, biking, leaping, etc.) may be tracked anciently using a camera attached to the chest or waist.

Whereas non-ambulation activities (such as brushing teeth, combing hair, eating, etc.) can be classified more efficiently using a wrist-worn sensor. This makes the use of more than one sensor in different locations of the body a good idea to improve both ambulation accuracy and non-ambulation activities. The sampling frequency of an accelerometer is also an important parameter. And studied the effect of different sampling frequencies on the accuracy of the classification. Maurer et al. found that the accuracy stabilized between 15 and 20Hz and not significantly improving with higher sampling rates. The sampling frequency was studied from an energy-saving perspective. They found that there is a trade specific to activity between consumed energy and classification accuracy based on sampling frequency.

So, based on the type of activity, they introduced a smart adaptive method that changes the sampling frequency in real-time. HAR has two main approaches: threshold-based and machine-based learning. Threshold-based HAR schemes do not require any training procedures but may be used to distinguish fairly small numbers of activities compared to systems that are based on machine learning techniques. The majority of HAR systems use supervised classification algorithms in order to classify the relevant activities, as shown in Table 1, making use of the ability of learning algorithms to detect and discriminate between different hidden patterns of activities.

Objective Function Based on Discriminant Feature

In objective value calculation, firstly, the Euclidean distance metric is used for calculating the distance between two samples in the n-dimensional feature space. If x and y are two points with n-dimensional features, the distance calculation can be formulated as follows:

dx,y=(x1−y1)2+(x2−y2)2+…+(xn−yn)2−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−√

=∑i=1n(xi−yi)2−−−−−−−−−−−√.

Based on the distance between all samples of a class, the class median is determined. The class media is an i-th data sample, which has minimum accumulated distance satisfying i=argmin {Dicumulated} with other samples in the class. The calculation of the cumulated distance of a data sample is formulated as follows:

Dicumulated=∑j=1n∥xi−yj∥2,i=1,2,…,n.

Approach:

Machine learning as a science is directly linked to quantitative statistics, and gaining historical experience in statistics is valuable for recognizing and manipulating algorithms of machine learning. Correlation is an interaction measure between two variables that are either not described as dependent or independent. To determine the association with a dependent variable, Regression allows for estimation capacities when the independent variable is known. Machine learning methods are constantly developed. Below are some of the more frequent ones.

Conclusion:

The research presents smartwatch sensor data based on the hybrid feature selection model, which robustly identifies various human activities. The data were obtained from inertial sensors (accelerometer and gyroscope) placed on the subject’s chest, i.e., people. The researchers recorded different human activities in a research lab. On the dataset, a minimum of 23 base functions is added. The sensors display a minimum of 138 heterogeneous elements. However, as a rule of thumb, not every feature contributes to activity recognition in the same way; rather, the classifier’s performance is degraded by extraneous features.

Thus, the proposed hybrid feature selection method containing the approaches to filter and wrapper has played an important role in selecting optimal functions. Using the SVM to classify human behaviors, the chosen apps are used for validation checking. The suggested method indicates an overall classification efficiency of 96.81 percent utilizing optimized apps, which is around a 6 percent better output increase without the selection of features. The model being suggested outperforms other state-of-the-art projects.

Cite this page