Research On Human Activity Recognition Based On Smart Watch

Activity recognition has important applications in motion monitoring,human-computer interaction,healthy life guidance and monitoring of the elderly and children,which nowadays has been playing an increasingly important role in people's daily life.However,traditional methods of activity recognition have some difficulties that can not be over-come.First of all,traditional methods need to collect a large number of training data to build the model,which waste much time and energy.Secondly,when we use the model to recognize a new user's motion,it would often lead to low recognise accuracy because of the user's deviations.In recent years,smart wristbands,smart watches gradually become popular and we will carry out our research on activity recognition based on their built-in inertial sensors.Our first work was research on human activity recognition based on arm pos-ture.We treat arm's dynamic movement as a series of static arm posture and we use the gravitational acceleration to depict arm posture.On the basis of above work,we can effectively recognize human's complex motions and achieve good scalability.Our second work is wearable approach for logic cognition-based activity sensing.Because the local coordinate system of the smart watch can rotate constantly when the human body moves,which makes the original data inconsistent,we establish a consistent lo-cal coordinate system of human body;Secondly,we extract the angle profile between the human arm and the local coordinate system of human body from the sensor data,and then describe the action of the human body through the angle profile;Finally,in order to tolerate individual differences,we decompose human actions into sequences of consecutive meta-activities,so that the loss or increase of a particular meta-activity will not lead to false identification.Through the above methods,we can realize user-independent activity recognition without requiring a large amount of training set data.In this paper,our main innovations are as follows:1.arm posture model We treat arm's dynamic movement as a series of static arm posture and we use the gravitational acceleration to depict arm posture.Because the gravity acceleration is relatively stable,we can effectively recognize human's complex motions and achieve good scalability.2.coordinate transformation of inertial sensors we set the vector which is par-allel to the front plane of the body to represent the first axis,and set the vector which is perpendicular to the front plane of the body to represent the second ax-is,and set the vector which is straight up to represent the third axis,and then we transfer the local coordinate system of inertial sensor to local coordinate system of human body through the global coordinate.Based on the consistent model of local coordinate system of human body,we can efficiently regular the original signal and improve the recognition accuracy.3.angle profile We treat arm as a rigid body and extract the angle between the arm direction and body's coordinate system.We future use the angle profile to depict the human's motion.Compared with the original data,the angle profile is more stable because it is only related to the trajectory of human's motion and has nothing to do with the speed and amplitude of the motion.4.meta-activity recognition We first extract the angle profile of the human arm.Then based on angle profile,we decompose complex activity into a sequence of meta-activities and recognize the original activity based on the sequence of meta-activities.By the fault tolerance of the sequence of meta-activities,we can effectively recognize the activity which may has lost or increased a particular meta-activity.