The manual processing method is mainly used for the detection, tracking and information editing of key objects in indoor scene video, which has the problems of low efficiency and low precision. In order to solve the problems, a new learning method of indoor scene semantic annotation based on texture information was proposed. Firstly, the optical flow method was used to obtain the motion information between video frames, and the key frame annotation and interframe motion information were used to initialize the annotation of non-key frames. Then, the image texture information constraint of non-key frames and its initialized annotation were used to construct an energy equation. Finally, the graph-cuts method was used for optimizing to obtain the solution of the energy equation, which was the non-key frame semantic annotation. The experimental results of the annotation accuracy and visual effects show that, compared with the motion estimation method and the model-based learning method, the proposed learning method of indoor scene semantic annotation based on texture information has the better effect. The proposed method can provide the reference for low-latency decision-making systems such as service robots, smart home and emergency response.

In the field of big data real-time analysis and computing, the importance of stream computing is constantly improved while the energy consumption of dealing with data on stream computing platform rises constantly. In order to solve the problems, an Energy-efficient Strategy for Threshold Control (ESTC) was proposed by changing the processing mode of node to data in stream computing. First of all, according to system load difference, the threshold of the work node was determined. Secondly, according to the threshold of the work node, the system data stream was randomly selected to determine the physical voltage of the adjustment system in different data processing situation. Finally, the system power was determined according to the different physical voltage. The experimental results and theoretical analysis show that in stream computing cluster consisting of 20 normal PCs, the system based on ESTC saves about 35.2% more energy than the original system. In addition, the ratio of performance and energy consumption under ESTC is 0.0803 tuple/(s·J), while the original system is 0.0698 tuple/(s·J). Therefore, the proposed ESTC can effectively reduce the energy consumption without affecting the system performance.

With the rapid development of mobile Internet, how to extract effective description information from a large number of mobile applications and then provide effective and accurate recommendation strategies for mobile users becomes urgent. At present, recommendation strategies are relatively traditional, and mostly recommend applications according to the single attribute, such as downloads, application name and application classification. In order to resolve the problem that the granularity of recommended applications is too coarse and the recommendation is not accurate, a mobile application similarity matrix construction method based on Latent Dirichlet Allocation (LDA) was proposed. Started from the application labels, a topic model distribution matrix of mobile applications was constructed, which was utilized to construct mobile application similarity matrix. Meanwhile, a method for converting the mobile application similarity matrix to the viable storage structure was also proposed. Extensive experiments demonstrate the feasibility of the proposed method, and the application similarity achieves 130 percent increasement by the proposed method compared with that by the existing 360 application market. The proposed method solves the problem that the recommended granularity is too coarse in the mobile application recommendation process, so that the recommendation result is more accurate.

Concerning the problem of data sparseness, cold start and low user experience of recommendation system, a video recommendation algorithm based on clustering and hierarchical model was proposed to improve the performance of recommendation system and user experience. Focusing on the user, similar users were obtained by analyzing Affiliation Propagation (AP) cluster, then historical data of online video of similar users was collected and a recommendation set of videos was geberated. Secondly, the user preference degree of a video was calculated and mapped into the tag weight of the video. Finally, a recommendation list of videos was generated by using analytic hierarchy model to calculate the ranking of user preference with videos. The experimental results on MovieLens Latest Dataset and YouTube video review text dataset show that the proposed algorithm has good performance in terms of Root-Mean-Square Error (RMSE) and the recommendation accuracy.

Concerning the problem of unbalanced load and incomplete comprehensive evaluation of nodes in big data stream processing platform, a dynamic load balancing strategy based on load awareness algorithm was proposed and applied to a data stream processing platform named Apache Flink. Firstly, the computational delay time of the nodes was obtained by using the depth-first search algorithm for the Directed Acyclic Graph (DAG) and regarded as the basis for evaluating the performance of the nodes, and the load balancing strategy was created. Secondly, the load migration technology for data stream was implemented based on the data block management strategy, and both the global and local load optimization was implemented through feedback. Finally, the feasibility of the algorithm was proved by evaluating its time-space complexity, meanwhile the influence of important parameters on the algorithm execution was discussed. The experimental results show that the proposed algorithm increases the efficiency of the task execution by optimizing the load sharing between nodes, and the task execution time is shortened by 6.51% averagely compared with the traditional load balancing strategy of Apache Flink.

Focusing on the issue that ZooKeeper cannot meet the requirement of low latency and quick recovery of RAMCloud, a Coordinator Election Strategy (CES) based on RAMCloud was proposed. First of all, according to the network environment of RAMCloud and factors of the coordinator itself, the performance indexes of coordinator were divided into two categories including individual indexes and coordinator indexes, and models for them were built separately. Next, the operation of RAMCloud was divided into error-free running period and data recovery period, their fitness functions were built separately, and then the two fitness functions were merged into a total fitness function according to time ratio. Lastly, on the basis of fitness value of RAMCloud Backup Coordinator (RBC), a new operator was proposed with randomness and the capacity of selecting an ideal target: CES would firstly eliminate poor-performing RBC by screening, as the range of choice was narrowed, CES would select the ultimate RBC from the collection of ideal coordinators by means of roulette. The experimental results showed that compared with other RBCs in the NS2 simulation environment, the coordinator selected by CES decreased latency by 19.35%; compared with ZooKeeper in the RAMCloud environment, the coordinator selected by CES reduced recovery time by 10.02%. In practical application of RAMCloud, the proposed CES can choose the coordinator with better performance, ensure the demand of low latency and quick recovery.

RAMCloud only supports the small object storage which is not larger than 1 MB. When the object which is larger than 1 MB needs to be stored in the RAMCloud cluster, it will be constrained by the object's size. So the big data objects can not be stored in the RAMCloud cluster. In order to resolve the storage limitation problem in RAMCloud, a parallel access strategy for big data objects based on RAMCloud was proposed. Firstly, the big data object was divided into several small data objects within 1 MB. Then the data summary was created in the client. The small data objects which were divided in the client were stored in RAMCloud cluster by the parallel access strategy. On the stage of reading, the data summary was firstly read, and then the small data objects were read in parallel from the RAMCloud cluster according to the data summary. Then the small data objects were merged into the big data object. The experimental results show that, the storage time of the proposed parallel access strategy for big data objects can reach 16 to 18 μs and the reading time can reach 6 to 7 μs without destroying the architecture of RAMCloud cluster. Under the InfiniBand network framework, the speedup of the proposed paralled strategy almost increases linearly, which can make the big data objects access rapidly and efficiently in microsecond level just like small data objects.

In order to solve the problem of low using rate, RAMCloud would change the positions of objects, which would cause the failure for Hash to localize the object, and the low efficiency of data search. On the other hand, since the needed data could not be positioned rapidly in the recovery process of the data, the returned segments from every single backup could not be organized perfectly. Due to such problems, RAMCloud Global Key (RGK) and binary index tree, as solutions, were proposed. RGK can be divided into three parts:positioned on master, on segment, and on object. The first two parts constituted Coordinator Index Key (CIK), which means in the recovery process, Coordinator Index Tree (CIT) could position the master of segments. The last two parts constituted Master Index Key (MIK), and Master Index Tree (MIT) could obtain objects quickly, even though the data was shifted the position in the memory. Compared with the traditional RAMCloud cluster, the time of obtaining objects can obviously reduce when the data throughput is increasing. Also, the idle time of coordinator and recombined time of log are both declining. The experimental results show that the global key with the support of the binary index tree can reduce the time of obtaining objects and recovering.