Evaluation method and device for automatic driving prediction module

Abstract

The invention discloses an evaluation method and device for an automatic driving prediction module, and relates to the technical field of automatic driving, and the method comprises the steps: carrying out the multi-level division of scene types, and screening and extracting a target scene segment from to-be-analyzed scene data according to an evaluation demand; determining an evaluation target and subordinate evaluation indexes thereof; constructing a judgment matrix according to the evaluation indexes, and obtaining a comprehensive weight value of each evaluation index to the evaluation target; acquiring historical scene fragment data of the same category as the target scene fragment in the database, and calculating a grey correlation coefficient between each evaluation index and the ideal reference sequence; and according to the comprehensive weight of each evaluation index and the grey correlation coefficient between each evaluation index and the ideal reference sequence, calculating to obtain a comprehensive evaluation value for evaluation of the prediction module. According to the method, the scene data containing multiple fragments can be automatically cut and extracted according to requirements, so that refined extraction and analysis of the scene are realized, and the reliability of an analysis result is effectively improved.

Description

technical field [0001] The invention relates to the technical field of automatic driving, in particular to an evaluation method and device for an automatic driving prediction module. Background technique [0002] The current evaluation and testing methods for prediction modules in the field of unmanned driving mainly have the following problems: [0003] 1. The scene analysis and extraction of the prediction module basically rely on manual screening and extraction, which is time-consuming and labor-intensive. In addition, for fragments with a relatively long time span and a combination of multiple scenes, they cannot be cut into individual pieces through automatic analysis and extraction. different scenes; [0004] 2. Scene extraction and module quantitative evaluation are independent of each other and not well integrated, resulting in a long overall cycle and low efficiency; [0005] 3. For the quantitative evaluation of the prediction module, only a few mainstream evalua...

AI Technical Summary

Problems solved by technology

[0003] 1. The scene analysis and extraction of the prediction module basically rely on manual screening and extraction, which is time-consuming and labor-intensive. In addition, for fragments with a relatively long time span and a combination of multiple scenes, they cannot be cut into individual pieces through automatic analysis and extraction. different scenes;

[0004] 2. Scene extraction and module quantitative evaluation are independent of each other and not well integrated, resulting in a long overall cycle and low efficiency;

[0005] 3. For the quantitative evaluation of the prediction module, only a few mainstream evaluation indicators are considered. At the same time, the evaluation results of the final module are mostly artificially given based on subjective experience or simply linearly superimposed on the aforementioned indicators, without sufficient consideration Differences between various indicators and differences between subjective experiences may lead to certain deviations or even problems in the output results;

[0006] 4. The evaluation and analysis of the prediction module does not include historical data into the overall analysis, and it is impossible to make a horizontal comparison between scenes of the same category. It is only judged from the analysis results of a single scene, and there may be certain evaluation deviations for some scenes

Images