Research And Application Of Video Event Detection Based On Deep Learning

Event detection from surveillance video is significant to the whole society,such as public safety,market analysis or production.But it is also a challenging task due to the characteristics of surveillance video such as large amount,low information density and various scenes.The detection process involves many technologies in image processing filed,such as image recognition,object detection,action recognition,image captioning and so forth.The algorithm has to consider both accuracy and speed with high availability,scalability and maintainability.In this paper,convolutional neural network(CNN)was used as feature extractor to carry out analysis on every single pixel and detected objects from images.In addition,3-D CNN was built to recognize actions from video,which has high accuracy and speed.Inspired by machine translation system,a CNN-RNN model was designed to form the map between image and natural language.This research also proposed a novel video event detection process combing fine grained and coarse grained analysis,which can not only extract key frames according to keywords,but also describe these image using natural language.Finally,a distributed embedded deep learning platform using storm was built to implement lightweight deep learning with low power consumption.All the proposed methods were evaluated and optimized according to experiments,demonstrating that these methods are suitable for real time video processing with high accuracy.