Research On SLAM Algorithm Based On Improved Mask-RCNN

Simultaneous Localization And Mapping(SLAM)refers to the localization and mapping of mobile robots by acquiring information about the surrounding environment through various sensors on board without a priori information about the external environment.Probabilistic SLAM algorithms are not suitable for large scale use due to the performance bottleneck of their own algorithms and the high cost of sensors.The visionbased SLAM algorithm uses vision sensors to capture environmental image information,solves the robot’s pose information based on image feature extraction and matching,and builds the corresponding trajectory map.With the development of artificial intelligence,visual SLAM algorithms incorporating deep learning can extract object feature information in images more finely and have become a popular research direction in recent years.In this thesis,the SLAM algorithm based on Mask-RCNN is used as the research direction to address the problem that the SLAM algorithm based on Mask-RCNN does not have high accuracy in image feature extraction and matching due to the influence of light and shadow factors brought by the occlusion and the problem that the feature extraction is less or cannot accurately identify the occluded object due to the influence of the occluded object.The proposed visual SLAM algorithm based on Probabilistic Mismatching Removal Algorithm reduces the influence of environmental changes such as light and shadows;the proposed visual SLAM algorithm based on Mixed Attention Mask-RCNN enhances the recognition and feature extraction ability of the occluded objects and improves the algorithm’s localization accuracy and composition accuracy.The algorithm is susceptible to changes in light and shadows when occlusion occurs in dynamic scenes,and changes in the pixel values of feature points in images lead to more false matches,larger errors in pose estimation,and the constructed trajectory maps are shifted or even cannot be built.In this thesis,a visual SLAM algorithm based on Probabilistic Mismatching Removal Algorithm is used to establish an inter-image pixel coordinate system model to calculate the inter-pixel distances of matched pairs of neighbouring image frames for positional resolution.The algorithm is able to enhance the matching ability of key frames,improve the matching accuracy,and ultimately reduce the error of the algorithm in constructing the trajectory map by removing the matching point pairs that do not satisfy the set threshold.The problem that the algorithm extracts fewer or fails to recognize the occluded objects when the occlusion occurs in dynamic scenes,resulting in the reduction or even absence of key features in the image.In this thesis,a visual SLAM algorithm based on Mixed Attention Mask-RCNN is used to improve the Mask-RCNN instance segmentation algorithm by fusing two soft attention mechanisms,channel and space.The Mask-RCNN algorithm backbone network architecture is adjusted to increase the weight of the occluded features,enhance the recognition ability and feature extraction ability of the occluded features,avoid the weakening or loss of the occluded features in the layer-by-layer calculation of the neural network,provide richer image semantic information for the closed-loop detection,and make the algorithm further improve the localization accuracy and build a more accurate trajectory map.In this thesis,the improved algorithm is validated with multiple sequences from the KITTI public dataset and TUM public dataset,and the corresponding hardware experimental platform is built to test the algorithm in real scenes.The experimental results show that the proposed visual SLAM algorithm based on probabilistic de-mismatching effectively reduces the sensitivity of the original algorithm to pixel changes caused by changes in light and shadows in dynamic occlusion scenes,significantly improves the algorithm’s ability to extract and match features on the collected image information,and obtains more accurate trajectory maps;the proposed visual SLAM algorithm incorporating the hybrid attention Mask-RCNN.By increasing the weight assigned to the occluded object,the original algorithm improves the recognition and feature extraction ability of the occluded object,reduces the trajectory drift error,and improves the localization ability of the mobile robot and the composition ability of the environment.