Lightweight visual identification method applied to small service robot

Abstract

The invention belongs to the field of intelligent robots, and discloses a lightweight visual identification method applied to a small service robot, the method comprises the following steps: carrying out image content acquisition and relative position acquisition by using a camera of the robot; performing data label marking on the obtained image; exporting the marked data labels, and storing picture data used for training; running and obtaining a trained visual identification model; storing the trained model file in the corresponding model file directory; running the model file, acquiring the image through the camera, and performing real-time target recognition; sending the target recognition result to a corresponding motion control module of the robot in real time, and realizing closed-loop control of the robot; and after debugging and script writing of the robot, controlling the robot to move. The complexity of the model is reduced, and the method is suitable for small-size and high-flexibility service robots.

Description

Technical field: [0001] The invention belongs to the field of intelligent robots, and more specifically relates to a lightweight visual recognition method applied to small service robots. Background technique: [0002] At present, yolov5 is an open source vision-based target recognition neural network technology with good effects. Most of the existing technical solutions using yolov5 are deployed on non-movable high-computing GPU computing platforms. The single-frame multi-classification function of yolov5 Improvements are made to realize simultaneous detection and recognition of multiple types of targets in the same frame of image. Since the visual recognition algorithm model is generally large and the depth of the model is large, it is not suitable for lightweight computing devices with low computing power. And the application scene is relatively single, in most cases there is only one target to be recognized in the image, so the calculation speed of the algorithm and the...

AI Technical Summary

Problems solved by technology

It solves the problem that the current visual recognition algorithm model is generally large and difficult to apply to lightweight devices with low computing power. It optimizes the scene of using CPU for model training and running the model, and reduces the classification by reducing the number of target classifications. The complexity of the network; the fuzzy kernel reduces the impact of dynamic blur on the model recognition effect during the high-speed movement of the robot

Images