Research And Implementation Of A Deep Learning Based Critical Event Spotting And Analysis System

In recent years,with the development of deep learning techniques,a series of breakthroughs have been made in the field of computer vision.At the same time,mobile phones,tablets and other mobile filming devices have become widely available,generating huge amounts of video data.How to automatically understand and analyse these video data has become a hot direction in the field of computer vision.Event spotting is a branch of this direction.Event spotting was proposed because researchers found that traditional action time detection has inherent errors,and that human perception of time range is rather vague,which leads to some bias in the annotation of the time range of the same action by annotators.The researchers therefore proposed the task of event discovery,which overcomes this bias by anchoring events to specific points in time.The thesis begins with an introduction to the relevant techniques involved in event spotting,with a particular focus on the two-stream network,as well as an explanation of the datasets.This thesis presents TwoStreamNet,a deep learning based event spotting algorithm,which uses a two-stream model architecture with a mixture of Convolution Neural Network(CNN)and Recurrent Neural Network(RNN);and is trained on standard datasets.The algorithm was tested on standard datasets and compared with the baseline model to validate its advancement and effectiveness.Based on the completion of the core algorithm,this paper designs and implements a deep learning-based critical event spotting and analysis system,completes the requirements analysis,outline design,detailed design and system tests.The deep learning-based event spotting algorithm TwoStreamNet has achieved significant improvement over the baseline network on both the GolfDB dataset and SoccerNet dataset;the key event spotting and analysis system has been fully tested and validated in real-world use.