Research On Key Technologies Of Monocular Visual Odometry Based On DSP

Monocular visual odometry has the advantage of simple structure, high flexibility and low cost which has recently applied for mobile robot, simultaneous localization and mapping, virtual city design etc. The whole system is based on personal computer in the concept of tradition, but the demerits of PC itself, including of large volume, weak stability and integration, bring severe restriction in the further application. Taking this into consideration, we present a new method, realizing monocular visual odometry in the DSP Embedded platform, which can improve the investigation and application by solving the problem properly.Firstly, the methods of localization were reviewed, and the theoris and limitations of each method were analyzed. The development state of monocular visual odometry was summarized, and a method of monocular visual odometry based on embedded system was brought forward.Secondly, conventional algorithms of feature tracking, feature matching and estimation of transform matrix was studied on PC platform. Comparison of feature tracking algorithms: Harris, SIFT and SURF was done, and a new method based on Harris and SURF was raised; The relationship between SURF parameters (octave number, octave layer number, threshold for determinant of Hessian matrix) and the localization accuracy and frequency was analyzed, suitable parameters were chosen and a methord of threshold for determinant of Hessian matrix dynamic setting was brought forward, based on the optimization of SURF parameters, time for feature tracking and matching was great reduced; Characters of feature matching in monocular visual odometry was analyzed, and a method of feature matching based on trace predication was brought forward, which improve the efficiency and correct rate.Thirdly, monocular visual odometry system based on PC platform was build. On this system, the effectiveness of improved methods was verified. And a long distance (506m, closed path) localization experiment was done based on the solution of feature tracking by SURF, feature matching by exhaustive search based on trace predication, similarity transformation matrix estimation by RANSAC. Experiment demonstrates that the accuracy of mileage is 0.2%, the accuracy of course bearing is 2.8%, and the localization frequency is up to 16fps.Finally, monocular visual odometry based on DSP embedded system was built. By comprehensive consideration of the localization accuracy, localization frequency, difficulty of programming and development cycle, the chosen solution is that feature tracking by Harris, feature matching by rotational invariance NCC based on trace predication, similarity transformation matrix estimation by RANSAC. A preliminary experiment by calculating the movement parameters between two adjacent images was done to verify the feasibility of the built system. Experiment shows that the embedded system can give a good result: translation motion accuracy is 0.99%, and rotation accuracy is 3.30%.