An End-to-End Unsupervised Scenario Passable Region Cognition and Understanding Approach

Abstract

The invention discloses an end-to-end unsupervised scene road area determination method. By constructing a road position prior probability distribution map and directly adding it to the convolution layer as a feature map of a detection network, a fusion position prior is constructed. The convolutional network framework of the feature, and then combine the full convolutional network and U-NET to construct a deep network architecture-UC-FCN network, and use the constructed prior probability distribution map of the passable area as a part of the deep network architecture-UC-FCN network A feature map is mapped to generate a UC-FCN-L network; the passable area is detected based on the vanishing point detection method, and the obtained detection result is used as the true value of the training data set to train the UC-FCN-L network, and the used The deep network model for the extraction of drivable areas solves the problem of difficult labeling of drivable areas. It has strong applicability, can work stably in various road environments, and has good real-time performance. This method has high detection accuracy, adaptability, and real-time The performance and robustness are good, and the method is simple and effective.

Description

technical field [0001] The invention belongs to the technical field of traffic control, and in particular relates to an end-to-end self-monitoring scene passable area cognition and understanding method based on a video data set. Background technique [0002] With the development of society, cars have become an irreplaceable means of transportation for human daily life. However, its security problems are becoming more and more prominent. The "Global Road Safety Status Report" pointed out that the number of deaths caused by traffic accidents is as high as 1.24 million per year, and the main causes of accidents are driver negligence and fatigue driving. In order to alleviate this situation, the development of automobile intelligent technology is particularly important , in the research of automatic driving and advanced assisted driving based on computer vision, the real-time cognition and understanding of the drivable area in front of the vehicle is an essential link. The driv...

AI Technical Summary

Problems solved by technology

At present, intelligent cars mainly combine radar and cameras to recognize and understand the driving area. However, radar (lidar, millimeter wave radar, ultrasonic radar) usually has high cost, high power consumption and is prone to mutual interference.

[0003] The vision-based drivable area cognition and understanding method is mainly based on the road surface color, road model, road surface texture features, etc. to obtain the basic structural features of the road surface, and further obtain the vanishing point, road edge line, and the basic direction of the road (direction) through these features. walking, left turn, right turn, sharp left turn, sharp right turn) and other potential information, use the traditional segmentation extraction method for the final extraction of the drivable area for these features, but this method of using traditional segmentation is often not ideal. Some traffic participants such as vehicles and pedestrians may be extracted into the drivable area, causing adverse effects on the next step of the smart car.

Images