Hierarchical self organizing map and focusing inspection strategy for mobile robot novelty detection
Novelty detection is a process of recognizing changes based on learned knowledge. In this research, a novelty detection system was implemented on a mobile robot with an array of sonar sensors for surveillance application. In order to perform novelty detection, a map that stores normal information w...
Saved in:
| 总结: | Novelty detection is a process of recognizing changes based on learned knowledge. In this research, a novelty detection system was implemented on a mobile robot with an array of sonar sensors for surveillance application. In order to perform novelty detection, a
map that stores normal information with respect to any particular robot pose in an environment is required. The map is needed to detect changes and determine the position of novel event. The challenges of mobile novelty detection system are that the false
positive rate is usually high whereas the true positive rate is usually low due to mapping
and monitoring problems. During mapping, errors due to robot localization and sensor measurement can reduce the quality of the map built. However, available methods in mapping assume perfect localization, hence error in localization is not taken into account
in the process of mapping. During monitoring, inspection interval that is too small will consume a lot of time and energy but if the interval is too big, novelty could be missed, hence lower the true positive detection. On top of that, low true positive detection is also caused by the low reliability of sonar sensor measurement. Thus, the objective of this thesis is to utilize mobile novelty detection system by developing a mapping and monitoring strategy that has low false positive detection, high true positive detection and able to estimate the position of a novelty. This thesis proposed two methods regarding to mapping and monitoring process; a hierarchical Self Organizing Map (SOM) and a Focusing Inspection Strategy (FIS). Unlike other mapping methods, hierarchical SOM also consider localization error when associating the normal information with respect to the robot pose. FIS is a multi resolution monitoring strategy which works by changing the frequency of measurement depending on the detection of anomaly. In this thesis, two models were considered; a step (FS) and linear (FL) resolution models. The hierarchical SOM was validated by using simulation and experimentation of the inspection in environment with normal and novel event. False positive rate is measured to determine the map performance. The results show that hierarchical SOM is able to map the normal condition of the environment very well. The inspection results show the false positive rate occurred less than 0.1 at the higher sensitivity setting of 0.9 in either normal or novel condition. The performance of FIS was investigated by using experimentation of the inspection of novel objects of different sizes. The results show that by changing the frequency of measurement using the FS and FL models, the number of true positive detection increases up to 80% when compared to inspection with fix measurement frequency. FIS also reduced the error of position estimation by about 8.8% and 10.9% each for FS and FL and maintained the false positive rate lower than 0.1. |
|---|