A non-contact weighing method

Abstract

The invention discloses a non-contact weighing method, which comprises: S1: preprocessing the data to be predicted to be sent into the neural network; S2: sending the preprocessed data to be predicted into the neural network to perform convolution feature Extraction; S3: Output the prediction result of the data to be predicted after the calculation of the neural network, wherein, step S3 adds a regression layer in the last part of the neural network, and makes the regression layer parallel with the classification layer and the positioning frame coordinate output layer respectively, realizing three output in parallel. The present invention uses an end-to-end deep neural network structure to process the three tasks of target positioning, identification and specific attribute value prediction, which simplifies the task of non-contact weight measurement under messy background and harsh environment conditions, and can significantly Improve the efficiency of weighing.

Description

technical field [0001] The invention relates to the technical field of data analysis, in particular to a non-contact weighing method. Background technique [0002] Target detection and positioning is the premise of three-dimensional target measurement. The existing target detection and positioning method and system based on depth data use the depth data collected by the depth data sensor to obtain a depth map, construct the background model of the depth map, and obtain the depth map from the perspective of the depth map. The target foreground area of ​​the target, the target foreground area is initially segmented, the obtained target foreground area is mapped to the top-down perspective, and the target foreground area under the top-down perspective is re-segmented according to the position difference of the target in the horizontal space to extract the depth of the foreground area Features and training to obtain the target detection model, and obtain the target detection res...

AI Technical Summary

Problems solved by technology

[0003] However, feature extraction is performed separately from the target detection model, which requires a large amount of computation and consumes a lot of memory. Under complex background conditions, if there are many noise points, it is difficult to segment a very clean target foreground. If the background modeling result is not ideal, the target will be There are small holes, which affect the accuracy of target detection

Images