In order to solve the problem of load imbalance of Distributed Hash Table(DHT) cloud storage system, this paper proposes a load balancing algorithm based on Node Dynamic Forward(NDF). Through overload nodes forward, this algorithm can reduce the node’s partition, and thereby reduce the storage node load. At the same time, copy the corresponding data to the third successor node, to ensure the stable number of copies the data in the system. In the NDF implementation process, it only needs a simple coordination between overload node and the third successor node. Therefore, multiple overloaded nodes can concurrently do load transfer, which is suitable for deployment in a large cluster. The functional test in a minor cluster of 10 nodes verifies that the NDF has good ability of load balancing. The performance test in a massive cluster of 5 000 nodes demonstrates the total system load does not exceed 60%, and compared with the virtual node algorithm, NDF algorithm can improve the load transfer cost by more than 30%.
null
Aiming at incomprehensive description of 3D model of Silhouette descriptor(SIL) and Density-Based Framework(DBF), this paper proposes a new 3D model retrieval algorithm: Density-based Contour(DBC). It characterizes a 3D object using multivariate probability functions of the object’s 2D contours’ features. Two models can be compared by the similarity of their 2D contours’ probability functions. The new algorithm performs better than SIL on describing contours of the model and it also shows stronger resistance to noise than DBF. The retrieval performance on PSB shows that DBC has a higher retrieval accuracy comparing to other traditional state-of-the-art 3D model retrieval algorithms.
For the problem that the tracker is hard to be resumed when particle filtering fails to track the target, this paper introduces a method that combines particle filtering with online learning. It uses the validated result of particle filtering as positive sample to update the training set. It uses random ferns as classifier to detect object. When there is a big difference between two results, the particle filter will be reinitialized. Two bit binary pattern is used as the online learning feature. It is easy to be computed, and has invariance to illumination and scale. Experimental result proves that this method has better tracking result than particle filtering and it can track the sheltered and disappeared target.