Method, device and system for deeply analyzing traffic scene

Abstract

The invention discloses a method and a device for deeply analyzing a traffic scene. The method comprises the steps of using a data set of original images in a plurality of traffic scene databases and road areas corresponding to the original images as training samples; through a Laplacian Pyramid transform mode, respectively resizing the original images into different scales, and inputting into neural networks respectively corresponding to the different scales, wherein each neural network is composed of a convolutional neural network part and a deconvolutional neural network part; outputting a one-dimensional array having the same pixel with the original image through a fully connected layer connected with each neural network, and restoring into a result image having the same size with the original image, wherein different types of roads are marked in the result image; processing the result image by a preset standard to restore the segmentation result of the roads; and inputting the image to be detected into the successfully trained neural network, and thereby obtaining the result image in which the road segmentation is completed and corresponding to the image to be detected. According to the method provided by the invention, the accuracy of analyzing the traffic scene can be improved.

Description

technical field [0001] The invention belongs to the technical field of data analysis, and in particular relates to a traffic scene depth analysis method, device and system. Background technique [0002] The traffic scene is extremely complex for car driving. The method of analyzing the camera sampling can efficiently and quickly evaluate the current scene. When applied to the automatic driving system and assisted driving system, one of the most important functions is to accurately analyze the scene image Middle road area information. [0003] In the existing technology, some use four-line lidar to detect roads, but this detection method has extremely high requirements on road edges. In roads without obvious boundaries, the road detection effect is not good and road data in changing traffic scenes Sampling is not universal. Some use statistical learning methods to fit the features of adjacent pixels. As long as the training samples of different types of roads are added, the...

AI Technical Summary

Problems solved by technology

[0003] In the existing technology, some use four-line lidar to detect roads, but this detection method has extremely high requirements on road edges. In roads without obvious boundaries, the road detection effect is not good and road data in changing traffic scenes Sampling is not universal

Some use statistical learning methods to fit the features of adjacent pixels. As long as the training samples of different types of roads are added, the road diversity problem can also be solved. However, using the maximum flow / minimum cut algorithm to obtain the road area, because the weight estimation will be Because of the distortion and unclearness of the actual road samples, the correct parameters cannot be obtained through the weight estimator well. In some particularly sensitive scenes caused by lighting and other reasons, this method also loses the ability to identify roads.

In the above existing technologies, because the traffic scenes are actually very diverse, they do not have the ability to adapt to different road conditions, and the recognition accuracy of the traffic scene road conditions is not high

Images