Study Of Object Detection Based On Infrared Imaging Video

Object detection is a challenging task in the field of computer vision.Unlike the traditional RGB three-channel video camera,the Thermal Infrared(TIR)camera has the advantage of being able to operate in completely dark environment and insensitive to light and shadow changes.It can provide noisy low-resolution images and is mainly used to capture targets relative to colder background.Infrared video imaging has great potential in military and security fields.Compared with the traditional RGB three-channel video research,the current studies on infrared video object detection are relatively few,and the related technology needs to be improved urgently.Aiming at the characteristics of thermal infrared video,such as complex noise,blurred picture quality,single channel and little image information,this thesis proposes an infrared video object detection model based on rectified optical flow algorithm using deep neural network technology,and uses super-resolution model to enhance the detection of small targets.The innovative work of this thesis is summarized as follows:1.Aiming at the problem of huge noise and zero drift in the normalization of the infrared video optical flow,this thesis proposes a non-linear optical flow rectifying algorithm to compute the optical flow such that all stationary targets can maintain the same optical flow value after optical flow normalization,and an efficient rectifying threshold that can maintain the discrimination degree of object motion information effectively after the optical flow rectification.2.This thesis improves the image-based single-stage object detection systems by combining the rectified optical flow image and the original video frame image with appearance information as their inputs.The performance of the existing main single-stage object detection methods are improved,especially the detection performance of the moving target lacking appearance information.Different from the traditional moving object detection models,the improved detection framework can also efficiently detect stationary targets,overcoming the shortcomings of moving object detection methods such as inter-frame difference method and achieving the detection of moving and stationary targets simultaneously.3.To deal with the fact that infrared video images contain less appearance information,this thesis proposes a super-resolution model based on deep neural network to expand and enhance the video image information.The experimental results show that the super-resolution video is more suitable to improve the detection of small targets.4.To deal with the shortcoming that only one target is labeled in the each video frame of the original VOT-TIR2016 data set,this thesis expands the data set label effectively to make it suitable for multi-target detection,which lays a solid foundation for the further study on the object detection based on infrared video.