Research And Implementation Of Tire Detection Technology Based On Weakly Supervised Deep Network

Fire is one of the most common catastrophic events caused by natural or human factor.The occurrence of fire not only can cause huge loss of property and heavy casualties,but also cause social confusion and destroy ecological balance.Therefore,it is becoming more and more important to detect the fire accurately and in real time in order to reduce its harm.The existing methods of fire detection are divided into two main types:fire detection based on the traditional image processing and fire detection based on the depth network.Because of the different levels of the flame,the image features of the corresponding images have great differences,and it’s limited to extract features as well as make classification manually.Therefore,this paper adopts the method of fire detection based on the weakly supervised deep network.(1)Propose a video fire detection method based on the weakly supervised two-stream ordinal regression deep network.Combined with the representation ability of the deep network,we formalize the fire detection to an ordinal regression problem with multiple class and propose a fire detection model based on the two-stream ordinal regression deep network,which fuse the motion information and visual information of key frames effectively.For evaluation,we conduct experiments on our data set and three publicly available benchmarks,the results in all the cases reveal that the proposed method improves accuracy significantly.(2)Build a large scale fire video database.Because the large scale database ImageNet is not suitable for fire detection,the small scale of the public data sets currently are not suitable for training a deep network.Thus,we first build an outdoor large scale fire video database with multi-scenes,which is called FireData.There are 700 video sequences annotated by human in the database,which are divided into 5 grades according to the proportion of the flame area in an image.(3)Design and develop a real-time video fire detection system.Based on C++ and OpenCV source library,a real-time video fire monitoring system is implemented in the operating system of Windows,which used the fire detection model proposed this paper.