| With the widespread deployment of large-scale network surveillance,vehicle retrieval has become a challenging task in intelligent transportation systems.Vehicle retrieval can accurately retrieve the vehicle and related information matching the target identity from the mass surveillance video data,which has a wide range of application value.Due to the similarity of vehicles themselves and the complex and varied surveillance scenes,existing vehicle retrieval algorithms rely on complex network models to mine and extract more discriminative vehicle image features,which leads to a decrease in retrieval efficiency.The mainstream vehicle retrieval systems only support image or vehicle appearance attribute retrieval,which cannot meet users’ diverse retrieval needs in complex surveillance scenarios.Besides,constrained by factors such as data size,feature matching algorithms,and hardware performance,vehicle retrieval systems have the problem of slow retrieval speed.This topic is from a joint project between Beijing University of Posts and Telecommunications and the Telecommunications Joint Laboratory.The topic is the research and system implementation of vehicle retrieval algorithm oriented to surveillance video,aiming to conduct research on the problems and challenges existing in vehicle retrieval algorithms and systems.The main research content and innovative achievements of this topic are as follows:1.In view of the problem of the complex network model of existing vehicle retrieval algorithms,this study proposes a vehicle retrieval algorithm based on multi-task learning,including a general multi-task learning framework and multi-task retrieval strategy.The multi-task learning framework adopts camera angle retrieval and vehicle orientation retrieval as auxiliary tasks to guide vehicle feature extraction based on multi-task learning,thereby effectively improving the performance of model without introducing too many model parameters.The multi-task retrieval strategy utilizes camera angle features and vehicle orientation features extracted from auxiliary tasks to eliminate bias caused by background and viewpoint changes,further solving intra-class differences caused by camera angle and vehicle orientation.On the VeRi-776 vehicle data set in surveillance scenes,the average accuracy of the algorithm proposed in this study is improved by 3.3%relative to the benchmark algorithm,and all indicators are better than the current state-of-the-art vehicle retrieval algorithm.2.In view of the problem of single function and slow retrieval speed of existing vehicle retrieval systems,this study,based on the intelligent transportation analysis platform,designs and develops a vehicle retrieval system oriented to surveillance video with multiple retrieval modes,including image-based retrieval,target attribute-based retrieval,and natural language description-based retrieval.Different from traditional vehicle retrieval systems,this study designs and implements a vehicle retrieval mode based on natural language description,which supports using dynamic attributes described in natural language to retrieve vehicles(such as vehicle movement and its interaction with the environment),thus meeting users’ diverse needs.Besides,this study introduces the Elasticsearch search engine for feature similarity search,effectively improving the system’s retrieval efficiency.The functional and performance tests of the system on real surveillance videos show that the system can meet the expected functional requirements,accuracy requirements,and retrieval time requirements. |
