Real-time Detection Method Of Moving Object Based On Infrared And Visible Light

Moving object detection is one of the important research directions of computer vision.It has been widely used in intelligent video surveillance,automatic industrial control,military defense and other fields.With the continuous development of moving target detection technology,the requirements for detection are increasingly high,including real-time detection,all-weather detection and detection in outdoor environment.Traditional moving target detection methods have many problems of false detection and low detection efficiency,which can not meet the needs of practical application.Therefore,this thesis carries out indepth research on moving target detection methods,proposes a moving target real-time detection method based on infrared and visible light,and develops a set of principle prototype.Specific research contents and results are as follows:1)A real-time moving target detection method based on deep learning and pipeline filtering algorithm is proposed.Firstly,the deep learning network is trained by using the characteristics of the target to be detected.Then a target detection algorithm based on regression model is used to identify the target position and its category information in a single frame video image.Finally,according to the continuity of the moving target in time domain,the pseudo target is eliminated by the improved pipeline filtering algorithm.This method can quickly and accurately identify moving objects in video images.2)Developed a set of low cost,low power consumption,which can work in the outdoor environment,day and night real-time detection system of moving object.The system is composed of hardware structure and software program,including dual-spectrum(infrared and visible light)camera,upper computer,turntable,lithium battery,etc.The software flow of the system includes initialization,video image acquisition,object recognition in single frame image,false object elimination,early warning and so on.The system program runs on the NVIDIA Ge Force GT 1030 graphics card and can achieve a processing speed of 25FPS(limited by the camera's acquisition frame rate).The total power consumption of the system is about 65 W.It is powered by lithium battery and can work in outdoor environment.Dual spectral camera can be used for 24 hours monitoring.The camera rotates with the turntable,which can realize 360° realtime monitoring without dead Angle.3)Outdoor experiments are carried out to verify the effectiveness of the real-time detection method proposed in this thesis.Before the experiment,COCO data set was used to train visible deep learning network model,and KAIST data set was used to train infrared deep learning network model.During the experiment,the dual-spectral camera was used to collect video and image data,and 5 groups of visible video and 4 groups of infrared video were obtained as experimental data.The moving pedestrians in the video were taken as detection objects,and the system software was used for detection.When a moving object was detected,an early warning signal was sent out.The experimental results show that the correct alarm rate is 82.37% and the false alarm rate is 1.20% in visible video.In the infrared video,the correct warning rate is 76.92% and the false alarm rate is 5.77%.