An Effective Activity Recognition Framework Of Objects' Behaviors In Real-world Circumstances

With the rapid development of current technology,activity recognition related technologies have made a considerable improvement over the past decade based on the advanced sensing technology.Nowadays,activity recognition is mostly considered as the technology that collects sensor data and uses machine learning models to classify the data collected,describing objects' actions.It has become much more important and common in daily life and brought benefits for human beings.However,the state-of-the-art researches mainly focus on developing better activity recognition algorithms to gain higher classification accuracy.These study are usually based on the laboratory circumstances,ignoring the problem of data effectiveness.In this paper we propose that the uncertainty during data collecting in real-world circumstances can cause great consequences,missing value problem.Without solving the problem,effective activity recognition is unreachable.In this paper,we study the missing value problem of activity recognition systems deployed in real-world circumstances.We propose missing value problem and summarize the problem as blackouts challenge and resource constraints challenge.To handle these challenges,we carefully study the data effectiveness issue in real-world circumstances and impute the blackouts caused by the ineffective data or other reasons.We impute the missing value by mining data characteristics to find similar data segments for reference,providing effective data for activity recognition.Based on the algorithm above,we propose an effective recognition framework of objects' behaviors which could work in real-world circumstances.We demonstrate the effectiveness of our proposed framework in human activity recognition circumstance and unmannedaerial vehicle activity recognition circumstance.Considering the different data characteristic under different circumstances,we realize each sub-module and combine them forming a runnable system.What's more,we collect two different datasets in the two different circumstances and design experiments carefully to verify the usefulness of our algorithm and system.In all,the contribution of this paper includes:1.We propose the missing value problem for activity recognition systems in real-world circumstance,and consider blackouts problem as the main concern.Then we model and analyze the challenging blackouts problems and point out the resource constraint problem can't be ignored.2.We propose an effective missing value imputation algorithm for blackouts based on time series data's similarity and smoothness.This algorithm performs well in both datasets and we construct an effective activity recognition framework EARFBS for real-world circumstances with this algorithm.In this way,we get effective and complete data stream for continuous activity recognition.3.We realize our framework as runnable system in smartphone-based human activity recognition circumstance and unmanned aerial vehicle activity recognition circumstance.And we collect two datasets under both circumstances and evaluate our system,demonstrating the effectiveness of our algorithm and framework.