Monocular vision obstacle avoidance method based on deep learning

Abstract

The invention provides a method for avoiding an obstacle based on a deep double-Q network countermeasure architecture. The method comprises the steps that a monocular vision RGB image is adopted, anda corresponding depth image is obtained; based on a countermeasure network and double-Q network mechanism, a model is trained in a simulator, and knowledge leant from a simulation test can be seamlessly transferred into a new scene in the real word; a machine learns how to avoid the obstacle on the simulator, and the deep information forecasting can be conducted even in an extremely noisy RGB image. According to the method, in combination with the double-current countermeasure network, monocular vision obstacle avoidance is conducted, the end-to-end high-speed learning of the obstacle avoidance task is achieved with the limited computing resources based on the double-Q network by the adoption of the countermeasure network architecture and can be directly transferred into the real robot completely, complex modeling and parameter adjustment of a traditional path planner are avoided, the performance can be improved greatly, and the training speed is increased greatly; and in addition, a variety of robot operating environment information is provided by a monocular camera, the cost is low, the weight is low, and the method is applicable to various platforms.

Description

technical field [0001] The present invention relates to the field of visual obstacle avoidance, in particular to a monocular visual obstacle avoidance method based on deep learning. Background technique [0002] Deep learning has shown great promise in robotics and computer vision. Deep learning-based path planning that guides learning how to avoid collisions is becoming increasingly popular. When mobile robots work in the real world, one of the basic capabilities they need is to be able to avoid obstacles due to various situations. Commonly used in UAV, aerospace, military reconnaissance, robotics, navigation planning and other fields. In particular, obstacle avoidance technology navigates flying robots in complex forest environments; as robots begin to enter factories, warehouses, hotels, shopping malls, restaurants and other environments, obstacle avoidance technology guides them to move. [0003] Obstacle avoidance typically utilizes ranging sensors such as laser scan...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the problems of high cost and low efficiency, the purpose of the present invention is to provide a monocular vision obstacle avoidance method based on deep learning, using monocular vision RGB images to obtain corresponding depth images, based on the confrontation network and double Q network mechanism, in the simulation The model is trained in the server, and the knowledge learned in the simulation test can be seamlessly transferred to the new scene in the real world

Images