Model training and obstacle detection method and device

Abstract

The invention discloses a model training and obstacle detection method and device, and the method comprises the steps: obtaining each frame of laser point cloud collected historically as a training sample, marking each training sample according to the obstacle information in each frame of laser point cloud, and, through each network layer of a to-be-trained obstacle detection model, and respectively determining a plurality of candidate detection frames of each obstacle in each frame of laser point cloud and a quality score of each candidate detection frame; then, clustering the candidate detection frames according to the quality scores of the candidate detection frames, and determining a plurality of high-quality detection frames; and finally, by taking minimization of the difference between the high-quality detection frame of each obstacle in each training sample and the position labeling frame of each obstacle in each training sample as a target, adjusting model parameters in the obstacle detection model. On the basis of the quality score of each candidate detection frame, clustering is carried out on each candidate detection frame, and the standard for judging the quality of the detection frames is dynamically adjusted, so that the generalization of the model for different scenes is improved.

Description

technical field [0001] The present application relates to the technical field of machine learning, in particular to a method and device for model training and obstacle detection. Background technique [0002] During the driving process, the unmanned driving equipment needs to detect the obstacle information of the surrounding obstacles in real time for obstacle avoidance driving. [0003] Among them, when performing obstacle detection, the laser point cloud in the surrounding environment can be collected by the laser radar device, and the collected laser point cloud can be input into the pre-trained obstacle detection model to determine the The position detection frame of each obstacle. [0004] At present, when training the obstacle detection model, several frames of laser point clouds collected in history can be obtained first, and the position labeling boxes of obstacles in each frame of laser point clouds can be determined. After that, for each frame of laser point clo...

AI Technical Summary

Problems solved by technology

[0006] However, since the laser point clouds used as training samples come from a variety of different scenes, when screening the pros and cons based on the intersection and comparison of detection frames, the same threshold standard is set for the laser point clouds collected in different scenes, resulting in the model targeting The generalization of different scenarios is poor, and the model output results are not accurate

For example, for two different scenarios with densely distributed obstacles and sparsely distributed obstacles, if the same screening criteria for good and bad detection frames are set, the distribution of good and bad detection frames in the two scenarios will be quite different.

Images