Research And Realization Of Monocular Thermal Visual Odometer Method In Outdoor Environment

Robot localization and navigation in the environment with degraded visual information is one of the current research hotspots of visual odometer.Based on visible light images,most visual odometry methods could only work stably in the environments with good lighting conditions.When applied at night or in the environments with fog,smoke,and dust,it is difficult to establish robust motion estimation from continuous visible images.Thermal images are used as the input of thermal visual odometer which can work robustly in the above-mentioned environment with degraded visual information.The thermal visual odometer can help the robot achieve all-weather localization and navigation in harsh environments,such as low light and smoke,which is of great significance to expand the application range of visual odometer with good application prospects.However,visual odometer relies on robust image features.Thermal images are not only characterized by the shortcomings of low resolution,low contrast,and lack of texture,but also the non-uniform noise,of which the inherent defects make the traditional visual odometry method unable to work robustly on thermal infrared images.To achieve precise positioning in the environment with degraded visual information,a monocular thermal visual odometry method was proposed in this thesis for outdoor environments,of which the research content is mainly divided into the following aspects(1)To solve the problem of low resolution and high noise of thermal image,taking the advantages of the characteristics of low contrast of thermal image and clear foreground and background,the ISODATA adaptive threshold algorithm with peak suppression was designed for the segmentation of the thermal image,and the segmentation result is regarded as the region of interest.The ROI-CLAHE enhancement algorithm of 14-bit thermal image was proposed to enhance the details of the foreground area while significantly suppressing the noise in the background area of the thermal image.(2)To solve the problem of sparse feature points and high mismatch rate due to the lack of thermal image texture,a feature point algorithm of joint thermal image segmentation boundary LK optical flow was proposed.The corner points of thermal image segmentation boundary were selected based on Harris response value,which were tracked by LK optical flow,and the outlier optical flow was removed according to the histogram of the optical flow direction.Then,the GMS-PROSAC algorithm was used to eliminate the excluded points according to the epipolar geometric constraint model.Regarding the information redundancy problem caused by the excessively concentrated distribution of feature points in the algorithm proposed in this thesis,the Grid-ANMS algorithm was proposed to uniformize the distribution of feature points.(3)In this thesis,a visual odometry method was extended to the thermal infrared imaging system,and the monocular thermal visual odometry system based on the outdoor environment was designed and implemented.The thermal visual odometer has a non-uniformity correction management module which ensures that the thermal infrared camera would not perform non-uniformity correction during large rotation movement,thereby avoiding the motion estimation failure due to the non-uniformity correction caused by the thermal infrared camera's violent rotation process.The system uses a thermal infrared camera as the visual sensor,which can effectively increase the number of feature points in the thermal image while reducing the mismatch rate of feature points.Besides,it can accurately solve the motion of the thermal infrared camera,and finally utilize the BA model to optimize the motion of the thermal infrared camera estimate.The experimental results of the comparison among ORB-SLAM2,Thermalvis and DSO in different scenes show that the monocular thermal visual odometry method proposed in this thesis could work robustly in the night environment and complete the accurate motion estimation of thermal infrared cameras,thereby realizing the all-day thermal visual odometer system that can work in low-light environment.