Research On Visual Localization Of Uavs Based On Motion State Keyframe Selection Strategy

As modern advanced equipment,UAVs play an important role in both civil and military fields.The high accuracy positioning of UAVs is the guarantee for their safe flight and the subsequent autonomous mission execution.However,the traditional GPS-based positioning solution cannot reliably position the UAV in complex scenes such as small and indoor areas.As limited by the weight and power of the UAV,the visual SLAM(Simultaneous Localization and Mapping)based positioning solution has started to become an option for the autonomous positioning of UAVs.It enables the high-precision positioning of drones through the onboard camera.However,the current visual SLAM suffers from the problem of easy drift or even tracking failure in a rotating or fast-moving scene.In this paper,we conduct a study on the anti-rotation of visual SLAM based on the keyframe strategy to address this problem.The main contributions of this paper are as follows:(1)A strong feature point screening method based on the survival time of features is proposed.It takes the number of frames that a feature point can be continuously tracked in a tracking thread as a metric,called the life value,and select out strong feature points as stable feature points depending on their life value.The selected strong features have reliable values and descriptor invariants,therefore,tracking them can reduce cumulative errors.In this paper,we build a visual odometer using such strong feature points.(2)A point feature SLAM algorithm based on motion-state-decided keyframes is proposed.Different from the traditional algorithms whose measurement of the image quality is pixel brightness and blur situation,this paper takes the number of strong feature points as the indicator of picture quality;different from the ORB-SLAM2 algorithm which generates keyframes based on time and space distance,this paper’s keyframe strategy takes into account the image quality and the motion state of the UAV and characterizes the intensity of rotational motion through the quantization of motion state based on the UAV’s poses(called the compound rotation amount).Besides,constraints such as the degree of overlap are added and form the strategy that is mainly based on strong feature points and secondarily based on the amount of composite rotation.The proposed method incorporates the visual odometry and keyframe strategies based on strong feature points and constructs a new weighted cost function at the back end for pose optimization.The comparison experiments showed that the absolute trajectory error of the algorithm can be reduced by 87%.(3)A point-line feature SLAM algorithm based on motion-state-decided keyframes is proposed,which is based on the improvement of PL-SLAM.All threads of this algorithm use point-line features.In terms of the keyframes selection,the strategy is constructed based on multiple constraints including the compound rotation amount and entropy function to evaluate the current rotational motion state and motion-estimation performance.The comparison experiments showed that the optimized keyframe strategy improves the localization accuracy of the algorithm,and the absolute trajectory error of the algorithm can be reduced by 66.4%.