A method and device for abnormal judgment based on vision

Abstract

The present application discloses a method and device for abnormality judgment based on vision, which is applied to an intelligent robot, wherein the method for abnormality judgment based on vision includes: acquiring a scene image; inputting the scene image into a spatial pyramid pooling-based The convolutional neural network, the feature map of the scene image is extracted by the convolutional neural network in the convolutional neural network based on spatial pyramid pooling, and the space in the convolutional neural network based on spatial pyramid pooling is used. The pyramid pooling network performs spatial pyramid pooling on the feature map to obtain a preset dimension feature vector; the preset dimension feature vector is input into the pre-trained anomaly detector, and according to the anomaly detector's The test results determine whether an abnormality has occurred.

Description

technical field [0001] The present application relates to the technical fields of image processing and intelligent robots, and in particular, to a method and device for abnormality judgment based on vision. Background technique [0002] In the process of performing tasks, intelligent robots often need to interact with the surrounding scenes of various types. This often requires the intelligent robot to prejudge whether the current scene conditions meet the conditions for executing subsequent actions, and whether it needs to stop the current task for exception handling, for example, by adjusting its position or posture to meet the conditions for executing subsequent actions. Abnormal judgment is the basis for the initial "intelligence" of intelligent robots. If the abnormal judgment method cannot quickly and accurately feed back the judgment results, it will cause the robot to run dry, and the robot interaction will be stiff and rigid, and the user experience will be poor. ...

AI Technical Summary

Problems solved by technology

However, this method requires the input of a fixed-size image, and the input image needs to be cropped or distorted. The cropped image may not contain the complete detection object, and the distorted image may be geometrically distorted, which will damage the accuracy of the judgment result.

In addition, the above method requires the detected object to occupy a relatively large proportion of the entire image, which limits the application scenarios of this method

Especially when it is necessary to use a small detected object as an abnormal and / or normal judgment standard in a large scene, because the detected object that provides key information accounts for a small proportion of the entire image , which often leads to the failure of the classification algorithm, and finally the correct judgment result cannot be obtained

Images