Novelty detection using a mobile robot : challenges and benefits
2017-01-15T23:00:55Z (GMT) by
Novelty detection is a method which highlights unusual data gathered from an environment. The use of novelty detection on a mobile system is an attractive idea. An important aspect of an intelligent robot is the ability to monitor changes as this is one of the important capabilities that are possessed by all biological systems. There are many applications that could benefit from this such as surveillance and inspection applications. However, there are also challenges that arise from implementing novelty detection on a mobile platform. One of them is the problem of mapping sensor measurements that are normally perceived (normal data) from the environment. Most conventional maps require some prior information about the environment to construct their structure, thus they are not readily adaptable to any environment. They also require a large amount of storage space and consequently require similar processing capability, and consume more power which as a whole constrains the design and size of the mobile system. This thesis presents an alternative mapping system for storing and learning normal data in the environment namely the flexible region mapping system. Its structure can change to accommodate the distribution of normal data. As a result, data are mapped where they are measured and according to the size of the affected area. This thesis also investigates approaches for reducing false positives and to estimate the position of anomalous objects by taking advantage of the system’s mobility. A close range inspection strategy has also been developed to demonstrate how an autonomous mobile robot could use the results of novelty detection to perform further investigation of anomalous objects. The work in this thesis has been targeted to be applicable to any mobile systems that could localized themselves, particularly those that have limited resources in terms of data storage, physical size, processing power and power supply. The solution of mapping normal sensor measurements that is demonstrated in this thesis is most suitable for structured environments but it could be extended to more complex environments. Experiments were conducted in an artificial L-shaped environment as well as in a real office corridor. A mobile robot that carries different types of sensors particularly a laser range finder, an anemometer, a temperature sensor, an ambient light sensor, a chemical concentration sensor and an electromagnetic radiation sensor was used. The results show that the flexible region map used as few as 0.7% and 3.3% of the storage space required for a conventional grid map and a perception based map. The map can autonomously accommodate to changes in the normal condition of the environment. The implementation of the false positive filter developed in this thesis reduces the false positive rate by up to 20% compared to the unfiltered novelty detection results, when using noisy sensors at the highest sensitivity settings. Apart from that, the close range inspection strategy is shown to be capable of achieving up to 100% close range inspection coverage near the vicinity of an anomaly.