Experiments in online expectation-based novelty-detection using 3D shape and colour perceptions for mobile robot inspection

Abstract Novelty detection is a very useful function for detecting abnormal data in any application. An expectation-based novelty-detection approach has been introduced that learns the dynamic relationship model among normal data in order to predict the next expected data. Most novelty-detection systems use an offline approach with a fixed structure, a system type that has limitations when the data count in the environment is unknown. A new expectation-based novelty-detection system features an online recurrent neural network approach that learns the data by inserting new nodes or deleting unused nodes from its structure. Generally, to detect novelties, a global novelty threshold is defined to filter out all input data as novel whenever the prediction error of the network exceeds the threshold. However, because a neural network cannot learn to predict all classes of input data perfectly, using a global novelty threshold leads to the misclassification of the insufficiently learned normal data as novel. To overcome this problem, the novelty-detection system has been improved to learn local novelty thresholds alongside its learning to predict expectations. The proposed algorithm is applied to an online novelty detection using colour and depth data obtained from a Kinect sensor on a mobile robot. The performance of the expected novelty detector and its limitations during experiments are analysed and shown. Furthermore, colour and depth data as inputs into the novelty filter are separately analysed and their contributions on the overall novelty detection highlighted. In conclusion, the performance of the novelty filter could further be improved by applying a better feature-selection technique to extract more interesting features from high-dimensional input data.