In my last blog post, Multi-sensor Data Fusion, I talked about the parts of a multi-sensor data fusion system and their roles. Using an admittedly over-simplistic example, I described how using different sources to add more information to the decision process can help make better and more precise decisions. In this post, I look at implementing these systems in a general way by examining two types of architecture and identifying their main strengths and weaknesses.
Multi-sensor systems can be divided into two types: centralized and distributed.
Figure 1: Centralized system architecture
Because each sensor is physically attached to the central processing unit, all the data streams from the sensors are received in parallel, making it easy to correlate the absolute time correspondence of the information. The ability to measure and compensate for the delay between each sensor’s information streams makes them easy to correlate during the data fusion stage.
Another strength is that the central processing unit can control each sensor. This means that it can quickly adjust sensor parameters in reaction to an event as well as detect any defects.
In a centralized system, the processing unit is often responsible for processing the data generated by every sensor in addition to performing the fusion of all the information and making decisions. This can result in a considerable processing load.
Centralized systems can typically only be used to cover relatively small areas. The fact that all the sensors are physically attached to the central processing unit implies proximity between them and the processor.
Figure 2: Distributed system architecture
Distributed systems can usually cover larger areas than their centralized counterparts, and the link between the sensors and the central processing unit can easily be adapted to the environment in which the system will be deployed.
Data processing directly performed on the sensors also helps to lighten the processing load on the central processing unit.
Distributed systems are usually more complex than centralized ones.
The data fusion on distributed systems must take into account that some data can be lost during communication with the remote sensors.
Also, the system must deal with the following problems related to the communication between the sensors and the central processing unit:
- A larger part of the data fusion process must be used to correlate the data of the different sensors in time since unpredictable time delays can occur.
- Data streams from one or more sensors can sometimes be received out of sequence.
In both approaches, the data fusion takes place on a central processing unit. There are other multi-sensor data fusion system architectures as well (for example, some “decentralized” systems implement partial data fusion centers at different nodes of the system). No system is the same and it is important to take into account every detail of the problem to solve before implementing such a system.