Automatic driving environment perception-oriented small sample in-loop learning system and method

Abstract

The invention relates to an automatic driving environment perception-oriented small sample in-loop learning system and method. Algorithm training is carried out on a target learning model based on a data training set to obtain a target detection algorithm; small sample training set training learning is carried out on the physical ECU module, and a target perception algorithm is obtained; target algorithm testing of the system is carried out through the obtained target perception algorithm. The system comprises a sample input module, an image acquisition module, an image display module, a physical ECU module and a result output module. According to the automatic driving environment perception-oriented small sample in-loop learning system and method, learning method based on small-scale datais provided, the arranged in-loop experiment simulation platform is matched with simulation software to solve the problem that large-scale data are difficult to obtain in real vehicle automatic driving environment perception algorithm development; the scene self-definition function is achieved, a large number of scene data which are difficult to achieve in actual situations are generated throughsimulation, and the experiment development safety is improved.

Description

technical field [0001] The invention relates to the field of intelligent driving, in particular to a small-sample-in-the-loop learning system and method for automatic driving environment perception. Background technique [0002] The autonomous driving environment perception system based on camera hardware is one of the most important components of the intelligent system of autonomous driving vehicles. The main functions of the autonomous driving intelligent system include: sensing, perception, decision-making, control, etc. In the development of autonomous vehicles, most of the technical difficulties and challenges come from the environmental perception and computer vision. [0003] At present, the solutions of autonomous driving environment perception technology are mainly based on computer vision and machine learning. Machine learning methods for computer vision are mainly based on deep learning. Deep learning is a data-driven machine learning method, which relies on a lar...

AI Technical Summary

Problems solved by technology

[0005] Taking an autonomous driving scene dataset as an example, there are 200 categories in total, among which the sample size of 10 high-frequency categories covers 45% of the overall dataset, and the sample size of 190 low-frequency categories covers 55% of the overall dataset. Most of them are low-frequency samples, and it is difficult to collect large-scale data for deep learning models for algorithm training

[0006] 2. Difficulty marking

[0007] Deep learning model training for autonomous driving environment perception often requires a data set of 100K scale. Large-scale data sets have the disadvantages of high cost and manual error in manual labeling, which leads to a lot of overhead for automatic driving companies in data labeling. , and cannot get accurate labeled data

[0008] 3. Insufficient data

[0009] In the large-scale sample data collection for algorithm model training, there are problems such as high data collection cost, long cycle, and difficulty in collecting dangerous scene data, which leads to insufficient sample data for autonomous driving companies.

Images