Visual SLAM closed-loop detection method based on depth neural network

Abstract

The invention discloses a visual SLAM closed-loop detection method based on a depth neural network, and the method comprises the following steps: training the network parameters of a linear decoder through a data set of a similar scene; carrying out the convolution processing of a collected image through the linear decoder; carrying out the dimension reduction of a high-dimension feature vector through a pooling method; Measuring the similarity of the features of the vector, obtained through training, through employing an inclined angle cosine function, and judging when to form a closed loop through setting a threshold value and combining the similarity of two scene images; and outputting a closed-loop detection accuracy recall rate curve and the detected closed loop for the subsequent SLAM mapping optimization. The method gives full consideration to the impact on the closed-loop detection accuracy and robustness from the descriptor of manual features, greatly improves the accuracy of an algorithm at the lower calculation cost, solves a problem of wrong closed-loop detection, facilitates the building of a more accurate map, and guarantees the consistency of generated maps.

Description

technical field [0001] The invention belongs to the technical field of image processing, and relates to a visual SLAM closed-loop detection method based on a deep neural network model. Background technique [0002] Mobile robots create maps based on sensor data in unknown environments and complete autonomous positioning, that is, simultaneous localization and mapping (SLAM) of the robot, which is the key to realizing autonomous mobile robots. The SLAM process generally includes several important modules, such as feature extraction and matching, data registration, closed-loop detection, and global optimization. Loop Closure detection refers to the mobile robot judging whether the current location has been visited, which is a key link in SLAM research. Accurately detecting the closed loop can effectively reduce the cumulative error of the robot pose estimation, which is conducive to building a more accurate map and ensuring the consistency of the generated map. Conversely, i...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the disadvantage that the traditional manual feature method is easily affected by the environment, the present invention uses the high-dimensional features obtained by the output layer of the neural network to describe the image, and uses these features to measure the image distance and establish a frame-to-frame feature association

Images