Topological graph-based intelligent navigation method and system

Abstract

The present invention discloses a topological graph-based intelligent navigation method and system, and relates to the field of computer technologies. A specific implementation manner of the method comprises: determining a navigable area view based on a constructed topological graph according to current position information; obtaining scene data, wherein the scene data comprises at least a scene image, a scene depth map, and a scene analytical diagram; and determining a behavior decision according to the navigable area view, the scene data, and a navigation model. The method is based on a multi-index navigation model, and the navigable area view is determined based on the constructed topological graph. Compared with a mapless mode, the method can be used to improve real-time navigation accuracy under the constraint of a GPS, and improve navigation robustness with incomplete dependence of the GPS.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a method and system for intelligent navigation based on topological graphs. Background technique [0002] Most of the existing autonomous driving technologies rely on high-definition maps for intelligent navigation. However, the process of obtaining high-definition maps is time-consuming and labor-intensive, and requires continuous maintenance and updates. In the existing technology, DeepMind researchers proposed a method based on deep reinforcement learning (2018Arxiv Learning to Navigate in Cities Without a Map) for intelligent navigation without maps in street view scenes. Deep reinforcement learning predicts the behavior of active entities (agents) by building a deep neural network model, and uses traditional reinforcement learning algorithms to independently train model parameters based on the rewards received in time series state (state) changes. The deep neural network ...

AI Technical Summary

Problems solved by technology

[0003] In the above-mentioned intelligent navigation method without map, its behavior only includes 5 kinds of decisions (rotate -67.5 degrees, -22.5 degrees, 22.5 degrees, 67.5 degrees and forward), the action decision is too single, it is difficult to deal with real driving

Secondly, this method relies entirely on images for intelligent navigation. Since images are affected by light, it is difficult to autonomously navigate under low light conditions such as night scenes.

In addition, the reward of this method is defined by a function of the distance from the target position. Generally speaking, a good navigation engine should be constrained by many aspects, including driving time, driving distance, and violations, etc. Therefore, it is difficult for this method to train multiple indicators. The real navigation model of

Finally, this method does not require any map, which will greatly expand the number of samples in the state space, the optimization process is not easy to converge, and the training is time-consuming and labor-intensive.

Images