Design And Implementation Of Video-based Product Placement Recognition System

With the development of the broadcasting and television industry and the improvement of people's quality of life,traditional continuous-segment advertisements have gradually been replaced by product placement due to the disadvantages of high investment costs,poor results,and bad audience experience.Product placement are highly concealed,flexible in delivery,and can affect the audience in a subtle way.However,the appearing time and position of the embedded ads are random,and traditional methods of manually selecting voice or image features for detection cannot accurately identify the embedded ads in the video.Therefore,this thesis introduces the object detection algorithm based on convolutional neural network into the detection of product placement,and designs and implements a video-oriented product placement recognition system,which provides a powerful tool for advertisement monitoring and delivery effect analysis.The main work done in this thesis is as follows:1.In view of the relevant background and current research at home and abroad,the system's research goal and content were clarified,that is,design and implement a product placement recognition system,which supports multiple object detection algorithms and various video streaming protocols as detection target.2.After the study of related theories and technologies such as convolutional neural network,Tensorflow,front-end framework Vue,Django framework and database,this thesis analyzed functional and non-functional requirements of the system,and then performed system outline and data storage design.After that,the thesis completed the division of system function modules and the design of front-end interface.3.3770 original product placement images of 10 brands were obtained by web crawler and frame extraction from TV program video.After manual annotation and data augmentation,21191 high-quality images were used as product placement datasets.4.The front-end and back-end functions of the product placement recognition system were designed and implemented in detail.Kafka is used as a message queue to decouple the detection algorithm from the video source to be detected.Under this specification,the access of YOLOv3 algorithm and UDP multicast over TS and HLS protocol video streams were completed.5.Perform detailed functional tests on each module of the system,and perform stress tests on core functions.The two-stage target detection networks of Cascade RCNN and Libra RCNN were set up,and the detection accuracy and speed were compared with YOLOv3,and the identification performance optimization schemes for different application scenarios were proposed.