| Bag-of-Features (BoF)-Based Human Motion Primitive Identification and Activity Recognition |
|---|
Project Description
The Bag-of-Features (BoF) model has been widely applied in many pattern recognition applications, such as text document classification, texture and object recognition and demonstrated impressive performance. In this project, we aim to explore the feasibility of applying a BoF-based framework for human motion primitives identification and activity recognition. The idea of motion primitives stems from human speech due to the similarity between human motion and speech signals. In speech recognition, sentences are first divided into isolated words. Each word is further divided into a sequence of phonemes. In English, there are about 50 phonemes shared by all the English words. Models are first built for each of these phonemes. These phoneme models then act as the building blocks to build words and sentences. Following the same idea, in this proejct, we build motion primitive-based activity model, where each activity is represented as a sequence of motion primitives which act as the smallest units to be modeled. We hope that these motion primitives capture the invariance aspects of the local features. Bof model is then built on top of these motion primitives for activity representation and recognition. Our experimental results validate the effectiveness of this framework for the task of human activity recognition. In addition, we have demonstrated that our statistical BoF framework can achieve a much better performance compared to the nonstatistical string-matching-based approach.
Publication
- Mi Zhang, and Alexander A. Sawchuk, "Motion Primitive-Based Human Activity Recognition Using a Bag-of-Features Approach", ACM SIGHIT International Health Informatics Symposium (IHI), Miami, Florida, USA, January 2012.  
- Mi Zhang, and Alexander A. Sawchuk, "A Bag-Of-Feature-Based Framework for Human Activity Representation and Recognition", International Conference on Ubiquitous Computing (Ubicomp) Workshop on Situation, Activity and Goal Awareness, Beijing, China, September 2011.  
| Learning and Recognizing Human Activity Manifolds |
|---|
Project Description
Manifold learning is an important technique for effective nonlinear dimensionality reduction in machine learning. In this project, we propose a framework based on manifold learning techniques that embeds the high-dimensional human activity signals into a low-dimensional subspace for compact representation and recognition. The idea of the manifold-based framework stems from the observation that the sensor signals of a subject performing certain activities are constrained by the physical body kinematics and the temporal constraints posed by the activity being performed. Given these constraints, it is expected that the sensor signals vary smoothly and lie on a low-dimensional manifold embedded in the high-dimensional input space. Moreover, these manifolds capture the intrinsic activity structures and act as trajectories to characterize different types of activities. This motivates the analysis of human activities in the low-dimensional manifold space rather than the high-dimensional input space.
Publication
- Mi Zhang, and Alexander A. Sawchuk, "Manifold Learning and Recognition of Human Activity Using Body-Area Sensors", International Conference on Machine Learning and Applications (ICMLA) Special Session on Machine Learning for Human Behavior Understanding and Assisted Living, Honolulu, Hawaii, USA, December 2011. 
| Feature Design and Hierarchical Feature Selection for Human Activity Recognition |
|---|
Project Description
It is well understood that high quality features are essential to improve the classification accuracy of any pattern recognition system. In this project, we focus on the problem of wearable sensor-based human activity recognition. The objective is to identify the most important features to differentiate various human activities. Specifically, we first design a set of new features (called physical features) based on the physical parameters of human motion to augment the commonly used statistical features. To systematically analyze the impact of the physical features on the performance of the recognition system, a single-layer feature selection framework is developed. Experimental results indicate that physical features are always among the top features selected by different feature selection methods and the recognition accuracy is generally improved to 90%, or 8% better than when only statistical features are used. Moreover, we show that the performance is further improved by 3.8% by extending the single-layer framework to a multi-layer framework which takes advantage of the inherent structure of human activities and performs feature selection and classification in a hierarchical manner.
Publication
- Mi Zhang, and Alexander A. Sawchuk, "A Feature Selection-Based Framework for Human Activity Recognition Using Wearable Multimodal Sensors", International Conference on Body Area Networks (BodyNets), Beijing, China, November 2011. 
| Fine-Grained Post-Stroke Motor Assessment Using Wearable Motion Sensors |
|---|
Project Description
Every year, approximately 795,000 people experience a new or recurrent stroke in the US, an event that is a leading cause of the motor deficit. Existing clinical studies demonstrate that motor function can be recovered through rehabilitation. To select the most appropriate rehabilitation programs, it is important to accurately assess the patient's current motor function. Traditionally, the assessment is based on the observations on the patient's motor behavior using standard clinical rating scales. However, this strategy has two drawbacks. First, since the assessment is based on the clinician's subjective judgments, the accuracy and consistency may vary significantly across clinicians. Second, the rating scales cannot record the details of the motor performance, thus failing to precisely evaluate the patient's progress during rehabilitation, To bridge the gaps, in this project, we develop a methodology for fine-grained assessments of post-stroke motion functionalities using wearable motion sensors. Our approach provides quantitative evaluations on motor function based on sensor signals and acts as a significant complement to the standard clinical rating scales.
Publication
- Mi Zhang, Belinda Lange, Chien-Yen Chang, and Alexander A. Sawchuk, "A Fine-Grained Analysis of Post-Stroke Motor Function Using Wearable Motion Sensors", AMA-IEEE Medical Technology Conference, Boston, Massachusetts, USA, October 2011. 
| Bayesian Networks-Based Automatic Fall Detection Using Context Information |
|---|
Project Description
In United States, Fall is responsible of 70% of accidental death in persons aged 75+. Therefore, an automatic fall detection system becomes a clear necessity. There are many existing fall detection systems having been developed. A majority of them adopt multi-stage thresholding technique where a fall activity is detected only if all the thresholds must be exceeded in a particular sequence. Time series analysis is another technique being used where a fall activity is detected if the waveform is matched with a pre-defined shape. However, no matter which technique is used, most of the systems focus on studying standalone fall activities, without taking the context information into consideration, which is not quite useful in the real-world settings. In this project, we are buidling a context-aware fall detection system based on Bayesian network. The original idea is that in general, people fall while they are walking or they just get up from beds. Our system is taking advantage of this context information to detect falls in a highly customized manner.
Publication
- Mi Zhang, and Alexander A. Sawchuk, "Context-Aware Fall Detection Using A Bayesian Network", International Conference on Ubiquitous Computing (Ubicomp) Workshop on Context-Awareness for Self-Managing Systems, Beijing, China, September 2011.  
| Coordinated Sampling in Body Area Sensor Network based on Markov Decision Process (MDP) |
|---|
Project Description
Power consumption is a critical issue in body area sensor network. One of the important sources of power consumption is data sampling at each sensor node. The higher the sampling rate is, the more power is consumed. In this project, we examine the feasibility of Markov Decision Process (MDP) as a static optimal decision policy for power control in body area sensor network. The MDP coordinated sampling policy can be learned based on the current human activity and the availability of energy at each sensor node. However, the space complexity of the representation of the learned MDP policy is exponential in the number of the sensor nodes. Since the learning of a small decision policy is key to deploying the model in small sensor nodes with limited memory, a compact representation of the decision policy is needed. Therefore, we build and compare the capability of the compact representation of the MDP policy by using different base supervised learners. The results show that unpruned decision trees and high confidence pruned decision trees provide the lowest error rate while the required node number of the decision tree is enough small to be stored in the sensors. Ensembles of lower-confidence trees are capable of perfect representation with only an order of magnitude increase in classifier size compared to individual pruned trees.
Publication
- Shuping Liu, and Mi Zhang, "Evaluation of Learning Algorithms for Optimal Policy Representation in Sensor-Network based Human Health Monitoring Systems", International Conference on Information, Communications and Signal Processing (ICICS), Macau, China, December 2009. 
| Reconfigurable Body Area Sensor Network for Motor Rehabilitation |
|---|
Project Description
An increasingly diverse set of sensing technologies is appearing in smart products and research projects. The goal of this project is to develop an interactive and intelligent wearable computing platform using various sensors for motion analysis, health monitoring, sports medicine, and motor rehabilitation/physical therapy. We have been building up a real-time multimodal patient monitoring Hardware/Software system based on SunSPOT technology. Many kinds of sensors, such as Bend Sensor, Flex/Force sensor, 3-Axis Accelerometers/Gyroscopes (IMU), IR/Ultrasonic rangers are being integrated onto the SunSPOT platforms. A fault-tolerant, dynamically configured Wireless Body Sensor Network has been set up to collect data from nodes attached to different parts of the human body. In-situ real-time feedback is provided based on sensor inputs to guide patients for proper movements. Meanwhile, the software installed on the PC not only supports startup calibration, displaying streaming data from multiple nodes, but also supports command-based remote reconfiguration on on-body nodes over the air. The data of each patient is streamed into a Weka machine learning engine, which is integrated in the backend PC to support real-time activity recognition.
Our prototype system also needs to be further improved in several aspects.
- Compared to mote-class systems (Tmote, MICA2), SunSPOT has a more powerful processing unit (32-bit ARM9 microprocessor) and a much larger data storage (4Mbit Flash). This makes in-network processing and local data storage possible, and therefore reduces radio usage so as to reduce power consumption.
- Most body area sensor systems are for individualized care. We are going to expand the system to enable multi-agent applications. Personal body area sensor system then can interact with the environment or other people within the inter-body network.
- Mi Zhang, and Alexander A. Sawchuk, "A Customizable Framework of Body Area Sensor Network for Rehabilitation", International Symposium on Applied Sciences in Biomedical and Communication Technologies (ISABEL), Bratislava, Slovakia, November 2009. 
Mobile Sensing and Image Processing
| OCRdroid: A Framework to Digitize Text on Smart Phones |
|---|
Project DescriptionAs demand grows for mobile phone applications, research in Optical Character Recognition (OCR), a technology well developed for scanned documents, is shifting focus to the recognition of text embedded in digital photographs. In this project, we present OCRdroid, a generic framework for developing OCR-based applications on mobile phones. We believe this mobile solution to extract information from physical world is a good match for future trend. OCRdroid combines a light-weight image preprocessing suite installed inside the mobile phone and an OCR engine connected to a backend server. We demonstrate the power and functionality of this framework by implementing two applications called PocketPal and PocketReader based on OCRdroid on HTC Android G1 mobile phone. Initial evaluations of these pilot experiments demonstrate the potential of using OCRdroid framework for realworld OCR-based mobile applications. For more information, please check the paper listed below and the project site with a demo video here.
Publication
- Mi Zhang, Anand Joshi, Ritesh Kadmawala, Karthik Dantu, Sameera Poduri, and Gaurav S. Sukhatme, "OCRdroid: A Framework to Digitize Text on Smart Phones", International Conference on Mobile Computing, Applications, and Services (MobiCASE), San Diego, California, USA, October 2009. 
| Mobile Labor Market: A General Platform for Mobile Community |
|---|
Project Description
Mobile Labor Market (MLM) is a project under “ Mobile Voices ”, which is a collaboration among Viterbi School of Engineering, the Annenberg School for Communication at University of Southern California, and the Institute of Popular Education of Southern California (IDEPSCA), a nonprofit institution serving low-income Latino immigrants in Southern California. This project aims to build a mobile platform for the research in mobile community. The platform is used by low-wage immigrants community in Los Angeles area, helping people without computers access to have greater participation in job huntings using their mobile phones. The GPRS-based multi-user software system adopts a Client-Server based 3-Tier architecture. All the information is password protected and ciphered for privacy. Data are updated in real time, and stored in the backend web-serviced database. A Nokia N95 smart phone is used acting as a basestation connected to the backend database. Local contractors and workers have corresponding programs installed on their mobile phones. A worker can upload and update his profile, search and apply for jobs which match his skills. A contractor can also upload and update his profile, post new jobs, search for job applicants, and pick up applicants for specific jobs. In addition, an evaluation system is introduced so that workers can be evaluated based on their performance. The prototype system triggers researchers’ interest at Nokia Research Center at Santa Monica, and now becomes the tool for Annenberg professors conducting their research.
Project Report (Unpublished Manuscript)
- Mi Zhang, Xiaoxing Chu, and Tutu Yu, "Mobile Labor Market: A General Platform for Mobile Community"  
