Monocular Infrared Image Depth Estimation Based On Unsupervised Learning

The acquisition of depth information plays an important role in many scenarios,such as inspection robots for obstacle avoidance,car assisted driving,and 3D reconstruction.Monocular depth estimation based on unsupervised learning can easily and quickly obtain the relative depth of all pixels in the image,which is also a hot research topic at present.However,the information contained in visible light images in low-light or nighttime environments is limited,and it is difficult for monocular depth estimation algorithms based on unsupervised learning to obtain accurate depth information.According to the characteristics of infrared images,this paper improves the monocular depth estimation algorithm based on unsupervised learning to make it suitable for infrared images,and solve the problem that depth information is difficult to obtain at night or in low light conditions.The main research contents are as follows:1.Make an infrared dataset.The monocular depth estimation algorithm based on unsupervised learning requires a large number of continuous frame images for network training,so about 33,000 infrared data sets are produced by themselves,with a resolution of1280*384.At present,the data set has been open sourced on the Arrow Open Source Infrared Platform.2.Improve the monocular depth estimation algorithm based on unsupervised learning to obtain the relative depth of infrared images.The improved algorithm adds a bidirectiona l feature pyramid network Bi FPN layer to the depth estimation network of the original algorithm model.The BIFPN structure can automatically learn and update the weights of different scale features in the final depth image,and can better integrate multi-scale feature information to solve the problem.The problem of blurred depth edges in infrared images.3.Propose an infrared image human-vehicle target absolute depth estimation algorithm to obtain the absolute distance of the human-vehicle target in the infrared image relative to the camera.The algorithm calculates the scale factor according to the principle of camera imaging geometry,and converts the relative depth into absolute depth through the scale factor.Then,combined with the target detection YOLOv5 algorithm,the human and vehicle targets in the infrared image are segmented,so as to obtain the absolute depth of the human and vehicle targets in the infrared image.The depth estimation method of infrared image human-vehicle target based on unsupervised learning proposed in this paper can be effectively applied to the fields of assisted driving of cars and obstacle avoidance of inspection robots.The average relative error of the algorithm within 20 meters is only 6.6%,which can meet the needs of obstacle avoidance.In addition,the inference speed of a single image on the Nvidia Xavier R&D chip only needs 40 ms,which can meet the real-time requirements.