Change-detection with limited situational awareness

Abstract

This work will focus on giving a perceptual system the ability to detect changes while maintaining its understanding of the environment. Most change detection systems can only perceive the change in the environment. They are not capable of processing the other objects in the environment. Nor are they capable of understanding what type of change they have just detected. This system aims to detect the difference between a novel object introduced to the environment and a known object moved within the environment while still segmenting the image. The image segmentation will use very high dimensional feature vectors. These will be obtained from multiple training images, and each percept will be given a specific label. The feature vectors will then be converted into sparse vectors and arranged in an approximate nearest neighbor (NN) search tree. The new image’s sparse vectors will scale the tree based on the Euclidian distances of the current sparse vector to the tree leaf nodes. The label from the leaf nodes will be selected as the representation of the percept in the new image. The novel objects will be detected based on a threshold distance from the leaf leave node. If this distance exceeds the threshold the object will be considered novel. The moved objects will be determined by a previously trained look up table (LUT). The LUT will hold a list of acceptable labels in for each pixel, and will be created from a series of training images. The results from the experiments show that this system is capable of learning the objects in an environment and understanding how the environment changes.