The specific objectives of this project are as follows:
As discussed throughout this report, the above objectives were answered through the acquisition and use of a commercial datalogger, the development of a substantial quantity of new software, design and implementation of some foot-pressure sensors and the testing of new belief network models produced for use with the new sensors and software.
The first project supervisor was Dr. Ann Nicholson from CSSE, who specialises in the field of Artificial Intelligence. Dr. Nicholson supervised previous fall detection projects and originally introduced the idea of using dynamic belief networks [2] to provide higher-level reasoning of ambulatory status using basic limb-sensor data (Section 3.2).
Dr. Ian Brown from the Monash Centre for Biomedical Engineering, who also co-supervised previous fall detection projects, is interested in the areas of rehabilitation technology, medical engineering and patient monitoring. Dr. Brown supervised the project's hardware and Medical aspects, as well as providing advice and requirements relating to ambulation monitoring and fall detection, especially in the elderly (Chapter 2). Through connections in the medical technology industry, Dr. Brown also acquired a datalogger for use in the project (Chapter 5).
Finite state machines are sequential digital logic circuits, consisting of digital inputs, small amounts of memory (usually single registers) and combinational state transition logic. At any point, the FSM should be in one of many possible finite `states' (hence the name) usually represented by a binary number stored in memory. When data is sampled by the FSM's inputs (usually at regular clocked time intervals) the combinational logic uses the inputs and current state to determine the next state of the FSM. As a consequence standard FSMs are generally first order Markovian. FSMs are generally very compact, fast and efficient.
This method, although useful in many digital situations where data is very reliable and deterministic, is difficult to apply successfully to environments such as ambulation monitoring, which produce continuous, potentially noisy and essentially analog sensor information. Additionally, any system attempting ambulation monitoring must deal with the physiological subtleties, unpredictable and complicated non-deterministic nature of the human bipedal gait (see Chapter 2).
Nicholson suggested that a class of uncertain reasoning systems called dynamic belief networks (DBNs - see section 3.2) developed in artificial intelligence would be more applicable to this domain. Based heavily in probability theory, DBNs can be used provide a degree of belief in the environment's (in this case a patient's) current state, as opposed to a definite state [2]. Such systems, although requiring more software and computational power, are generally more accurate and resistant to noise than purely deterministic systems, easier to alter or extend and in some cases incorporate `learning' facilities.
McGowan built a patient-mounted datalogger based on a Motorola HC11 micro-controller, which was used to collect and digitise analog sensor readings and store them into on-board memory. The readings were then down-loaded into a computer where DBN software processed the data to produce detection and prediction of falls. McGowan's project didn't incorporate radio telemetry as was done in the previous project, so relied on two distinct stages of recording and then processing. The rudimentary DBN software was built using existing software libraries written by Amy Holden during a summer scholarship project [6].
McInerney's project focused on testing and assessing the application of the newly developed commercial datalogger (shared with this project) in monitoring the ambulation and medical status of spinal injury patients.
Chapter 2 introduces the medical and physiological mechanisms behind walking and falling, and explains terms such as gait cycle and stepping strategy, which are referred to later in the report. Chapter 3 covers causal reasoning and Dynamic Belief Network (DBN) theory, which was used by the project to detect falls. An overview of the falls system's design is presented in Chapter 4, which introduces the hardware, software and their interaction, laying foundations for more detailed discussion in following chapters.
The first half of the project's hardware aspect, the Siesta datalogger, is introduced in Chapter 5, followed by details of the various sensors used with the datalogger to record ambulation information in Chapter 6. Implementation of the software developed during the project to process the ambulation data using DBN theory is explained in Chapter 7. Detailed documentation relating to the configuration and operation of the software is also provided in Appendix A).
Chapter 8 explains the structure and parameters of some DBN models tested via the software to monitor ambulation and detect falls. The results obtained from these models when used to process sensor recordings from different ambulation situations are presented and discussed in Chapter 9. Finally, Chapter 10 discusses the outcomes and achievements of the project, followed by the conclusion in Chapter 11.
The appendices cover technical details relating to design, implementation and operation of the project's software.
![\includegraphics[width=100mm]{figures/daviesFSM.eps}](img3.gif)