US 12,307,789 B2
Multi-task joint perception network model and detection method for traffic road surface information
Hai Wang, Zhenjiang (CN); Guirong Zhang, Zhenjiang (CN); Yingfeng Cai, Zhenjiang (CN); Long Chen, Zhenjiang (CN); Yicheng Li, Zhenjiang (CN); and Qingchao Liu, Zhenjiang (CN)
Assigned to JIANGSU UNIVERSITY, Zhenjiang (CN)
Appl. No. 18/575,391
Filed by JIANGSU UNIVERSITY, Zhenjiang (CN)
PCT Filed May 6, 2023, PCT No. PCT/CN2023/092501
§ 371(c)(1), (2) Date Dec. 29, 2023,
PCT Pub. No. WO2024/138993, PCT Pub. Date Jul. 4, 2024.
Claims priority of application No. 202211675099.6 (CN), filed on Dec. 26, 2022.
Prior Publication US 2024/0420487 A1, Dec. 19, 2024
Int. Cl. G06K 9/62 (2022.01); G06V 10/26 (2022.01); G06V 10/32 (2022.01); G06V 10/77 (2022.01); G06V 10/776 (2022.01); G06V 10/80 (2022.01); G06V 10/82 (2022.01); G06V 20/56 (2022.01)
CPC G06V 20/588 (2022.01) [G06V 10/26 (2022.01); G06V 10/32 (2022.01); G06V 10/7715 (2022.01); G06V 10/776 (2022.01); G06V 10/806 (2022.01); G06V 10/82 (2022.01)] 13 Claims
OG exemplary drawing
 
1. A multi-task joint perception method implementing a network model-for traffic road surface data, comprising:
preprocessing an input two-dimensional image to change a luminosity and a geometric shape of the image through data enhancement in a preprocessing process, by specific operations of adjusting a tone and saturation of the image and randomly rotating, zooming, translating, cutting, and overturning the image, to obtain a preprocessed image;
slicing the preprocessed image to obtain a sliced image, wherein information of H and W dimensions in the preprocessed image is concentrated into a channel space, such that the preprocessed image becomes a double downsampled image without an information loss, and then, the sliced image is transmitted;
downsampling the sliced image three times to extract image features;
receiving the image features to fuse the image features;
enhancing a receptive field of the model;
upsampling the fused image features three times, to obtain a restored image having an original size of the input two-dimensional image; and
performing an Add operation on a feature map obtained by first upsampling of the fused image features in a drivable area segmentation head and a feature map obtained by first upsampling of the fused image features in a lane line detection head, and input a feature map obtained after the Add operation into the lane line detection head for second upsampling.