CN-121999457-A - Automobile light control method and system based on deep learning

CN121999457ACN 121999457 ACN121999457 ACN 121999457ACN-121999457-A

Abstract

The invention provides an automobile light control method and system based on deep learning, and relates to the technical field of automobile lighting, wherein the method comprises the steps of obtaining public driving data; according to the disclosed driving data, constructing a night low-illumination image and a target annotation data set, constructing a night image enhancement network based on LE-net, inputting the night low-illumination image into the night image enhancement network, training the night image enhancement network, constructing a target detection network, inputting the target annotation data set into the target detection network, training the target detection network, acquiring a real-time image from a front camera, carrying out image enhancement on the real-time image through the night image enhancement network, carrying out target detection on the enhanced real-time image through the target detection network, carrying out glare risk assessment according to a target detection result to obtain a glare risk index, and carrying out switching control on a high beam lamp and a low beam lamp of a car lamp according to the glare risk index.

Inventors

TAN TING
WU SHULONG
XIE YONG
CHEN ZHENHUA
MA YULIN

Assignees

无锡机电高等职业技术学校
无锡商业职业技术学院
江苏慧诺德科技发展有限公司

Dates

Publication Date: 20260508
Application Date: 20260129

Claims (10)

1. The automobile light control method based on deep learning is characterized by comprising the following steps of: S1, acquiring public driving data; s2, constructing a night low-illumination image and a target annotation data set according to the public driving data; S3, constructing a night image enhancement network based on LE-net, inputting the night low-illumination image into the night image enhancement network, and training the night image enhancement network; S4, constructing a target detection network based on LCG-YOLOv-tiny; s5, acquiring a real-time image from a front camera of the automobile; s6, performing image enhancement on the real-time image through a trained night image enhancement network; s7, performing target detection on the enhanced real-time image through the target detection network; s8, carrying out glare risk assessment according to a target detection result to obtain a glare risk index; And S9, switching and controlling the high beam and the low beam of the car lamp according to the glare risk index.
2. The deep learning-based automobile light control method according to claim 1, wherein S2 specifically comprises: s201, acquiring a daytime image and a nighttime image from the public driving data; S202, darkening the daytime image through gamma conversion to obtain a darkened image; S203, performing contrast scaling operation on the darkened image to generate a low-illumination image; S204, aligning the gray level distribution of the low-illumination image and the night image through a histogram matching function to obtain a target image; s205, randomly selecting a partial region of the target image; s206, darkening the partial area through a partial shade distributed in Gaussian, so as to obtain the night low-illumination image; and S207, screening effective labels according to the distance of the high beam influence and the target area, and removing target frames outside the irradiation range of the vehicle lamp in the night low-illumination image to obtain the target label data set.
3. The deep learning-based automobile light control method according to claim 1, wherein the step S3 specifically comprises: s301, enhancing the night low-illumination image through the LE-net to obtain an enhanced image; S302, constructing a reconstruction error based on MSE according to the enhanced image and the daytime reference image; s303, calculating the peak signal-to-noise ratio of the enhanced image according to the reconstruction error; S304, calculating the structural similarity of the enhanced image and the daytime reference image; s305, constructing a comprehensive loss function of the night image enhancement network according to self-attention distillation, the peak signal-to-noise ratio and the structural similarity; S307, obtaining optimal night image enhancement network parameters by counter-propagating with the aim of minimizing the comprehensive loss function of the night image enhancement network; And S308, repeating the steps S301 to S307 until the maximum iteration times are reached, and outputting the optimal night image enhancement network parameters to obtain the night image enhancement network with the training completed.
4. The deep learning-based vehicle light control method according to claim 1, further comprising, after S4: And converting the training night image enhancement network and the training target detection network into a vehicle-mounted joint perception module and a light control module.
5. The vehicle light control method based on deep learning according to claim 4, wherein the conversion process of the vehicle-mounted combined sensing module and the light control module specifically comprises: building the vehicle-mounted joint perception module according to the night image enhancement network, the target detection network and the image acquisition module which are completed by training; Docking the vehicle-mounted joint sensing module with a vehicle body CAN bus and a lamplight executing mechanism; Mapping the car light command instruction output by the car-mounted joint sensing module into a car light state signal and an illumination adjusting signal; And sending the car light state signal and the illumination adjusting signal to the light control module through the car body CAN bus to finish the butt joint of the car body CAN bus, the car-mounted joint sensing module and the light control module.
6. The vehicle light control method based on deep learning according to claim 1, wherein the switching control is specifically to smoothly change the light brightness in a plurality of sampling periods by setting an illuminance gradual change curve.
7. The deep learning-based automobile light control method according to claim 1, wherein S9 specifically comprises: s901, generating a dynamic risk threshold according to the glare risk index; S902, calculating a far and near light switching threshold according to the evaluation index and the dynamic risk threshold, and determining a first threshold and a second threshold; S903, judging whether the maximum value of the glare risk index is larger than or equal to the first threshold value, if so, judging that the high glare risk exists, outputting a control instruction, switching the car lamp from the high beam lamp to the low beam lamp or reducing the illumination, and entering a step S904; s904, judging whether the maximum value of the glare risk index is smaller than or equal to the second threshold value, if so, recovering the car lamp to the high beam, otherwise, not switching the light of the car lamp.
8. The deep learning-based vehicle light control method of claim 7, wherein the evaluation index comprises vehicle speed, road type, and ambient brightness.
9. An automobile light control system based on deep learning is characterized by comprising a processor and a memory; The memory stores a program or instructions executable on the processor, which when executed by the processor, implement the steps of the deep learning based vehicle light control method of any one of claims 1 to 8.
10. A readable storage medium, wherein a program or instructions is stored on the readable storage medium, which when executed by a processor, implements the steps of the deep learning based vehicle light control method of any one of claims 1 to 8.

Description

Automobile light control method and system based on deep learning Technical Field The invention relates to the technical field of automobile lighting, in particular to an automobile light control method and system based on deep learning. Background In night or low-illumination driving scenes, the vehicle-mounted visual system faces a plurality of challenges, and the current mainstream car lamp control scheme is difficult to adapt to complex illumination conditions and dynamic traffic environments, so that target detection performance and driving safety are easy to influence. In the prior art, effective joint modeling and optimization are not available among night image enhancement, target detection and car lamp control, and real-time performance, detection precision and control robustness are difficult to be considered, so that a novel car lamp far and near light switching control method which is integrated with deep learning, is adaptive to a car-mounted edge computing platform and realizes perception-decision closed loop optimization is needed. The current mainstream car light control scheme is mainly based on simple image enhancement technologies such as fixed rule logic judgment or global histogram equalization, a multi-focus single-task model is researched by night image enhancement, deep learning methods such as GAN and CNN are partially adopted, the target detection field comprises two types of high-precision two-stage detectors such as FasterR-CNN and lightweight detectors such as YOLO series, and a car light control strategy is mainly based on a static threshold decision mechanism based on parameters such as ambient brightness and car speed. However, the prior art has the defects that the simple image enhancement technology is difficult to cope with complex illumination, the effects of detail preservation, noise suppression and the like are poor, meanwhile, the calculation cost of the two-stage detector is high, the false detection rate of missed detection of the lightweight detector under low illumination is high, the generalization capability is poor, in addition, the static threshold car lamp control strategy cannot be adapted to a dynamic traffic environment, the problems of switching lag and the like are easy to occur, and the intelligent level and the adaptability are insufficient due to the lack of a glare risk quantification assessment mechanism. Disclosure of Invention In view of the above shortcomings of the prior art, an object of an embodiment of the present invention is to provide an automobile light control method based on deep learning, which can solve the problems in the prior art that a simple image enhancement technology is difficult to cope with complex illumination, and effects such as detail retention and noise suppression are poor. Meanwhile, the two-stage detector has high calculation cost, and the lightweight detector has high false detection rate and poor generalization capability under low illumination. In addition, the static threshold car light control strategy cannot adapt to a dynamic traffic environment, the problems of switching hysteresis and the like are easy to occur, and the technical problems of insufficient intelligent level and adaptability due to lack of a glare risk quantitative evaluation mechanism are solved. In a first aspect of the embodiment of the present invention, an automobile light control method based on deep learning is provided, including: S1, acquiring public driving data. S2, constructing a night low-illumination image and a target annotation data set according to the public driving data. And S3, constructing a night image enhancement network based on the LE-net, inputting the night low-illumination image into the night image enhancement network, and training the night image enhancement network. And S4, constructing a target detection network based on LCG-YOLOv-tiny, inputting a target labeling dataset into the target detection network, and training the target detection network. S5, acquiring a real-time image from the front camera of the automobile. And S6, carrying out image enhancement on the real-time image through a night image enhancement network. And S7, carrying out target detection on the enhanced real-time image through a target detection network. And S8, carrying out glare risk assessment according to the target detection result to obtain a glare risk index. And S9, switching and controlling the high beam and the low beam of the car lamp according to the glare risk index. In a second aspect of the embodiment of the invention, an automobile light control system based on deep learning is provided, which comprises a processor and a memory. The memory stores programs or instructions executable on the processor which when executed by the processor implement the steps of the deep learning based automotive light control method as in the first aspect. In a third aspect of the embodiments of the present invention, a readable storage medium i