On-node Data Processing

From Capsil Wiki

When on-node sensor data processing is considered, the resource constraints of currently available low-power micro-processors have to be taken into consideration. Examples of previous work in this area include:

• A two layered model combining a Gaussian Mixture Model and a Markov Model was investigated in MIThril 2003. However, only the feature extraction and Gaussian class-conditional posterior code were implemented on the 200MIPS StrongArm processor of the [MIThril] system.
• An algorithm with low complexity was proposed in Sola et al. for ambulatory activity classification using a decision tree based technique to classify activities including resting, lying down, walking and running. The algorithm was implemented on a fixed-point Philips LPC2106 30MHz micro controller. The 32-bit ARM processor can be considered to be a relatively high power processor compared to micro controllers in typical wireless sensors. In addition to that, decision tree classifiers can be computationally intensive as logic operations are expensive processes. Re-training decision trees could also lead to a change in their structure, meaning that the node programming has to be changed.
• Several papers for on-line ECG detection, including: Akazawa et al, Park et al. (the wearable motherboard), Luprano et al. (Myheart project),and So and Chan (real-time ambulatory cardiac monitoring)
• An on-node classifier for an ear worn sensor that uses a Bayesian classifier with Gaussian class conditional densities in Lo et al. .
• An on-node detector of heart rate and SPO₂ in Wang et al. using a miniaturized ear-worn sensor.