Research On Real-Time Improvement Based On Visual SLAM

Simultaneous Localization and Mapping Technology is crucial to the realization of AR / VR,robot,unmanned aerial vehicle(UAV),and unmanned driving technologies.Due to the complex structure and large amount of calculation,the SLAM system has high requirements on hardware and poor real-time performance.With the growing trend of equipment miniaturization and low cost,how to optimize the calculation of SLAM system has become one of the hot topics in recent years.This paper focuses on the problem of large calculation amount of the visual odometer module and the back-end optimization module in the optimized visual SLAM system.Based on the existing visual SLAM system,the writer improved the feature extraction algorithm and the cumulative error filtering algorithm,proposing an improved algorithm for SIFT and IMM filter,which are respectively optimized from the front end and back end of the visual SLAM system,thus greatly improving the efficiency of the real-time visual SLAM system.The main research work and innovations of the thesis mainly include the following aspects:(1)Under complex conditions,the writer analyzed and compared several classic local image feature extraction methods through experiments,introduced their respective adaptability conditions,and explained their advantages and disadvantages.(2)To solve the problem that the traditional algorithm has low accuracy and low efficiency for feature point matching,the writer proposed an improved image matching method,that is,based on the scale-invariant characteristic transform(SIFT)algorithm,the extra calculation of unstable feature points can be reduced to significantly improve the matching efficiency by calculating the entropy difference of the neighborhood gray value to select stable feature points;the efficiency of screening out false matching can be effectively improved by optimizing the removal algorithm for false matching.The experimental results show that the improved SIFT algorithm basically retains the correct matching point pairs while the matching efficiency has been greatly improved.(3)To address the problem that the front-end visual odometer will inevitably produce error accumulation during the work,the writer proposed a back-end algorithm based on IMM filter.Combining the non-maneuvering model and the Singer model,the improved IMM algorithm can be well adapted to the strong maneuvering target by changing the maneuvering frequency and acceleration variance.The online experimental results of the standard data set show that compared with the traditional nonlinear optimization algorithm,the improved IMM algorithm can effectively improve the operating efficiency of the system,thus achieving a stable tracking of the target.