Research On Software And Hardware Codesign Of Monocular Visual Odometry

Many detection tasks of the future depend to a large extent on highly autonomous mobile robots equipped with cutting-edge scientific instruments and advanced navigation capabilities.In order to enable these mobile robots to explore larger areas,it is necessary to increase their navigation speed.Visual Odometry(VO),as part of mobile robots,has attracted more and more attention because of its reliable performance and a wide variety of applications.However,visual odometrys require complex computer vision algorithms to run on the CPU such as feature detection and so on,which requires longer execution times,resulting in relatively slower movements.Therefore,achieving high complexity algorithms on robots and achieving higher orders of magnitude has become a top priority.In order to solve this problem,this topic has carried out research on the software and hardware codesign of monocular visual odometry.Aiming at the calculation of each step of the monocular visual odometry based on the feature point method,the calculation task of the monocular visual odometr y is reasonably divided and mapped to the software and hardware of the system.The SURF feature point detection algorithm,the BRIEF point feature description algorithm and the feature point matching algorithm are implemented by using the FPGA resources on SOC,and the data is transmitted by VDMA.The ARM of the SOC is used to control the FPGA side and data scheduling.In order to achieve highspeed transmission between SOC and PC,a high-speed UDP/IP protocol stack core is designed by FPGA,and the matched feature points are transmitted to the PC through the core.Finally,the pose data is calculated on the PC to realize the pose estimation of the camera.In this paper,the verification system of the monocular visual odometry is designed and implemented according to the overall scheme,and the performance analysis and verification of each module of the whole system are carried out.Test results show that the Gigabit Ethernet transmission module can achieve a stable transmission bandwidth of up to 900 Mbps at a working frequency of 129 MHz.Using the monocular VO data set in the KITTI data set,the entire verification syst em was tested and found that the root mean square error of the pose estimation and the real trajectory was 0.5 m at the maximum and the minimum was 0.01 m.The rotation error is at most 0.515 deg/m and the minimum is 0.0110 deg/m.The translation error is up to 11.35% and the minimum is 1.38%.And the pose estimation output speed of every two frames reaches 33 Hz.The results show that the system design achieves the expected accuracy and meets the requirements of real-time processing.