The Abnormal Activity Recognition Of Human Based On The Wearable Sensor Network

Behavior detection is one of important research fields of ubiquitous computing, which might understand the status of human such as behavior, activity strength and energy consumption. With the development of sensor technique and low power wireless communication technique, behavior detection with wireless wearable technique extends its wide applications such as smart home, elder person monitoring, patient monitoring and rehabilitation training of athletes, etc..Currently, behavior detection with wearable technique focuses on pedometer, fall detetion and daily activities. For a special class of people such as prisoners and violent persons, little has been done on the abnormal behavior detection. For those people, if abnormal activity might be monitored, it will help to take appropriate actions in time and prevent further danger accidents when the abnormal activity is appearing. This thesis studies the abnormal behavior detection using wearable technique. It uses 3-axis accelerometer to monitor the activities such as walking and standing. The mean, variance, and standard variance in time domain are chosen as characteristics to detect the abnoral behaviors. The detailed work includes the following:(1) We study the detection accuracy and computation complexity of classification algorithms for abnomal behavior detection such as SVM(Support Vector Machine), KNN(K-Nearest Neighbors), DT(Decision Tree) and NB( Na?ve Bayes), and find that KNN performs best among them. Additionally, we discuss the performance of KNN by chosing the value of k in detail, the detection accuracy is more than 95% when k is 5, 7, 9 and 11.(2) To use with wearable equipment, we study the impact of training samples on the detection accuracy and compuation. We find that the number of training samples affects the detection accuracy little. For example, the detection accuracy for abnormal behavior decreases about 5%and 2%, when the number of training samples changes from 479 to 128 and 128 to 64 respectively. And using the ReliefF algorithm to calculate the weight of each feature to improve the KNN algorithm.(3) In wearable sensor platform-Shimmer, we verify the KNN algorithm, and find that in the wearable equipment, the algorithm needs to be optimized due to the constraint of compuation capacity and internal storage. For an instance, we compact the characteristic matrix consisting of mean, variance and standard variance from 9-dimension to 3-dimension, the accuracy of online activity recognition is about 60%.The results of this thesis will help the usage of wearable technique into the abnormal behavior detection.