Segmentation of moving objects using Multiple Background Model for industrial mobile robots
Taeho Kim, Kang-Hyun Jo · 2008
This paper investigates new approach for not only segmentation of moving objects but also generating background model from a camera on an unexpectedly moving mobile robot. The image sequence by moving camera is unstable. Therefore we have to estimate camera displacement for generating multiple background model. We estimate the camera displacement by similarities between two consecutive images and use correlation to calculate it in this paper. However, correlation process has heavy cost if we search whole image using small searching window. To reduce computational cost, firstly we detect strong vertical edges with selected regions which include detected edges from the current image. After we segment selected regions, the process chooses searching areas for each region to calculate similarity between consecutive images. The similarity generates displacement vectors using two center points of highly correlated regions for each selected region. The histogram of displacement vectors offers the camera displacement vector which has high density. Based on camera displacement vector, a pixel in previous image will match to displaced pixel in current image. Therefore the algorithm generates multiple background model (MBM) for each matched pixel by camera displacement vector. MBM process classifies each matched pixel to several clusters. Finally we eliminate clusters which have lower weight than threshold, and combine remained clusters for each pixel to generate multiple background model. Experimental results show that generated multiple background model and detected moving object under moving and shaking camera.