Optimization On Deep Learning-based Video Analytics For Multi-tenant

Large-scale video analytics is becoming particularly important for many applications such as surveillance,traffic control and online video search.Applying deep learning models(e.g.,object detector and face recognizer)to video data analytics is an effective method due to its high accuracy.At any time,there can be multiple queries from different users with similar concerns over the same video data.Taking into account that different queries often perform similar work,which brings many opportunities for data sharing.The performance can thereby be improved by reducing the redundant computation across queries.This thesis proposes a deep learning based multi-tenant sharing video analytics system called RDShare,tailored for deep learning-based video analytics by caching and reusing the computed inference results.It takes into account the data sharing of both the inter-queries and intra-query with the awareness of different accuracy required among queries.This thesis typically shows that 1)there is a large potential sharing possibility between queries.In particular,object detection results can be partially reused by object recognition,and vice versa;2)improving the accuracies of existing queries can increase the sharing opportunities for incoming queries dramatically,which however has a negative impact on the latency of existing queries.RDShare aims to maximize the data sharing between queries with no impact on a single query with the help of intra-query sharing and intra-query sharing.For intraquery sharing,it has a difference detector for determining the data reuse within the video of a query.For inter-query,it contains a query pruning component that prunes the pending queries with some overlaps and a temporal query sharing component that caches the inference results of existing queries and reuses the computed inference results for currently computing queries.Moreover,to save the storage space,this thesis proposes a compression approach that can reduce the storage of input video frames and cached output results significantly.Finally,this thesis experimentally evaluates RDShare and the experimental results validate the effectiveness of our approach.