SENSOR FUSION FRAMEWORK AND SIMULATION ON A TURTLEBOT ROBOTIC VEHICLE
by Shruti Gangadhar, 2017
2. SENSOR FUSION OVERVIEW
2.1 Sensor Data Acquisition
- 센서 종류에 따라 포맷이 다르다
2.2 Need for Sensor Fusion
- 여러 종류의 센서 데이터를 활용하면 정확도가 올라 간다.
2.3 Sensor Fusion Challenges
Data imperfection: Data from the sensors contain some amount of noise and imprecision. Data fusion algorithms should be able to take advantage of the redundant data to minimize the effects of such imperfections.
Outliers and spurious data: Ambiguity and inconsistencies in the environment that the sensors may not be able to distinguish, causing the measured data to be unreliable. Such data appear as outliers in the data set.
Conflicting data: If two sensors are offering conflicting data about an aspect under observation, the fusion algorithm should be able to handle such conflicts to avoid counter-intuitive results.
Data modality: The fusion process must take into consideration both qualitatively similar (homogeneous) and different (heterogeneous) sensor data.
Data correlation: When sensors are spatially distributed in a system, some sensor nodes are prone to external disturbances. This can bias the sensor readings and the fusion result may suffer from over/under confidence.
Data alignment: Data from various sensors must be brought to a common frame of reference before the fusion process. It deals with the calibration error induced by individual sensors.
Operational timing: The data used for fusion may be coming from sensors that span a vast area or from sensors that are generating data at different rates. Out-ofsequence arrival of data for fusion process can result in performance degradation especially in real time applications.
2.4 Categories of Sensor Fusion
Depending upon the sensor configuration, there are three main categories of sensor fusion: Complementary, Competitive and Co-operative [23].
A. Complementary
In this method, each sensor provides data about different aspects or attributes of the environment. By combining the data from each of the sensors we can arrive at a more global view of the environment or situation. Since there is no dependency between the sensors combining the data is relatively easy [23] [24].
B. Competitive
In this method, as the name suggests, several sensors measure the same or similar attributes. The data from several sensors is used to determine the overall value for the attribute under measurement. The measurements are taken independently and can also include measurements at different time instants for a single sensor. This method is useful in fault tolerant architectures to provide increased reliability of the measurement [23] [24].
C. Co-operative
When the data from two or more independent sensors in the system is required to derive information, then co-operative sensor networks are used since a sensor individually cannot give the required information regarding the environment. A common example is stereoscopic vision [23] [24]. Several other types of sensor networks exist such as corroborative, concordant, redundant etc [21]. Most of them are derived from the aforementioned sensor fusion categories.
2.5 Fusion Methodologies
Generally, sensor fusion framework is designed or chosen based on the application.
A. JDL Model
JDL stands for the US Joint Directors of Laboratories that was established under the guidance of Department of Defense and was proposed in 1985
B. Waterfall Fusion Process Model
The Waterfall fusion process model (WFFM) deals with the low-level processing of data and is shown in Figure 3
C. Dasarathy’s Classification
D. Category Theory Based Model by Kokar et al.
2.6 Sensor Fusion Topologies
A. Centralized
In this architecture, a single node handles the fusion process.
The sensors undergo preprocessing before they are sent to the central node for the fusion process to take place.
B. Decentralized
In this architecture, each of the sensor processes data at its node and there is no need for a global or central node. Since the information is processed individually at the node, it is used in applications that are large and widespread such as huge automated plants, spacecraft health monitoring etc [24]
C.Hybrid
This architecture is a combination of both centralized and distributed type.
When there are constraints on the system such as a requirement of less computational workload or limitations on the communication bandwidth, distributed scheme can be enabled.
Centralized fusion can be used when higher accuracy is necessary [24] [32].
2.7 Categories of Fusion Algorithms
Sensor fusion can be performed at various levels based on the condition and type of data.
In this context, there are following fusion stages:
- Signal level fusion
- Feature level fusion
- Decision level fusion
A. Signal Level Fusion
- data from multiple sources (sensors) are combined to obtain better quality data
가. 목적
목적 #1 : 동일한 센서(온도계 3개)에서 데이터를 수집하여 불확실성 제거
- Obtain a higher quality version of the input signals i.e. higher signal to noise ratio [33].
- Sensor measurements from several sensors which have same physical properties are combined to determine the parameter being measured, more accurately [21].
- This minimizes and sometimes eliminates any uncertainty or inaccurate predictions caused by measurements from faulty sensors, measurement noise and state noise.
- For instance, readings from multiple temperature sensors in close proximity in a given space can be used for this kind of fusion.
목적 #2 : 이기종 센서를 이용하여 새로운 feature 생성
- Obtain a feature or mid-level information about the system that a single measuring node cannot reveal.
- A feature is the first stage in understanding the state of the environment that helps the system in formulating a decision.
- Heterogeneous sensors are often employed for this process.
- For instance, signals from radar and images from camera are used in target recognition [25].
나. 방법론
- Weighted Averaging : Taking an average of the various sensor signals measuring a particular parameter of the environment
- Kalman Filter : common adaptive method of sensor fusion to remove redundancy in the system and to predict the state of the system
- Track-to-Track Fusion
- Neural Networks
B. Decision level fusion (=Symbol level fusion)
가. 목적
The decision level fusion combines several sub-decisions or features to yield a final or higher decision that can be used to take an action
Symbol could be an input decision
나. 방법론
- Dempster-Shafer Theory of Evidence
- Comparison of D-S and Bayesian Fusion