Distributed Visual Processing Based On interest Point Clustering

• Main Author: Bai, Xueyao
• Format: Others
• Language: English
• Published: KTH, Kommunikationsnät 2015
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-168013

In this master thesis project, we study the problem in Visual Sensor Networks in which only limited bandwidth is provided. The task is to search for ways to decrease the transmitting data on the camera side, and distribute the data to dif- ferent nodes. To do so, we extract the interest points on the camera side by using BRISK in- terest point detector, and we distribute the detected interest points into di erent number of processing node by implementing proposed clustering methods, namely, Number Based Clustering, K-Means Clustering and DBSCAN Clustering. Our results show it is useful to extract interest points on the camera side, which can reduce almost three quarters of data in the network. A step further, by imple- menting the clustering algorithms, we obtained the gain in overhead ratio, interest point imbalance and pixel processing load imbalance, respectively. Specically, the results show that none of the proposed clustering methods is better than oth- ers. Number Based Clustering can balance the processing load between di erent processing nodes perfectly, but performs bad in saving the bandwidth resources. K-Means Clustering performs middle in the evaluation while DBSCAN is great in saving the bandwidth resources but leads to a bad processing balance performance among the processing nodes.