Search

CN-121999468-A - Automatic driving obstacle detection method and system based on image recognition

CN121999468ACN 121999468 ACN121999468 ACN 121999468ACN-121999468-A

Abstract

The invention relates to the field of vehicle image processing, and particularly discloses an automatic driving obstacle detection method and system based on image recognition, wherein the method comprises the steps of detecting and segmenting a local degradation region in an original image to generate an accurate mask; the method comprises the steps of restoring the area based on a physical degradation model inverse process, generating a global restoration image through fusion, evaluating the reliability of the restoration area, generating a restoration reliability map to modulate depth characteristics to obtain an anti-restoration artifact attention characteristic map, detecting the characteristic maps of a current frame and a historical frame through a time sequence cooperative attention mechanism, and outputting a stable obstacle list, wherein the system comprises four modules with corresponding functions. According to the invention, by introducing a repair reliability evaluation and characteristic purification mechanism and combining with time sequence consistency verification, the deep problem of local repair artifact interference detection is systematically solved, and the accuracy and the robustness of obstacle detection under severe conditions such as rain, snow, mud and the like are remarkably improved.

Inventors

  • Zhang Wentaiyang

Assignees

  • 西安科技大学

Dates

Publication Date
20260508
Application Date
20260130

Claims (10)

  1. 1. The automatic driving obstacle detection method based on image recognition is characterized by comprising the following steps of: Detecting an instantaneous local degradation area in an original image of a current frame, and generating an accurate mask; reconstructing and fusing image content based on the accurate mask, and outputting a global repair image; Evaluating the reliability of a repair area in the global repair image, generating a repair reliability map, weighting image features according to the repair reliability map, and outputting an anti-repair artifact attention feature map; and performing obstacle detection by using a timing coordinated attention mechanism based on the anti-repair artifact attention feature maps of the current frame and the historical frame, and outputting an obstacle list.
  2. 2. The method for detecting an automatic driving obstacle based on image recognition according to claim 1, wherein the detecting an instantaneous local degradation region in an original image of a current frame, generating an accurate mask, specifically comprises: reading and caching an original image of a current frame in real time through at least one optical camera interface in the vehicle-mounted looking-around system; Identifying an instantaneous local degradation area caused by non-fixed attachments in the original image of the current frame; And (3) based on a point spread function physical degradation model of water drops and mud points, carrying out fuzzy kernel estimation on the identified candidate instantaneous local degradation areas, and removing false detection areas by combining gradient direction consistency verification in image gradient analysis to generate an accurate mask with accurate boundaries.
  3. 3. The image recognition-based automatic driving obstacle detection method according to claim 2, wherein the identifying of the instantaneous local degradation region caused by the non-fixed attachments in the original image specifically includes: comparing the pixel level difference of the original image of the current frame with that of a historical clear reference frame before a preset frame number, and identifying the instantaneous local degradation area which accords with the characteristic of non-fixed attachments by combining the change mode of the local gradient amplitude of the image; Wherein the change pattern comprises anisotropic attenuation characteristics of an edge diffusion gradient, and the instantaneous local degradation region comprises an image blurring or distortion region caused by water drops, splashed mud points or local lens contamination.
  4. 4. The method for detecting the automatic driving obstacle based on image recognition according to claim 2, wherein the image content reconstruction and fusion are performed based on the accurate mask, and a global repair image is output, specifically comprising: Expanding a preset pixel width outwards by taking the boundary of the area marked by the accurate mask as a reference, and cutting out a corresponding clear neighborhood image block which is not blocked from the original image of the current frame; based on the inverse process of the physical degradation model, reconstructing image content of the degradation area marked by the accurate mask to generate a local repair image; and carrying out smooth transition fusion on the undegraded region of the local repair image and the original image at the boundary of the accurate mask by using a poisson fusion algorithm, and outputting a global repair image.
  5. 5. The method for detecting an automatic driving obstacle based on image recognition according to claim 4, wherein the reconstructing of the image content of the degradation area identified by the precision mask based on the inverse process of the physical degradation model, to generate a local repair image, specifically comprises: and performing iterative optimization reconstruction on pixels in the precise mask region by using the inverse process of the physical degradation model and the blind deconvolution algorithm based on the point spread function estimation and taking the clear neighborhood image block as a priori constraint to generate a local repair image.
  6. 6. The method for detecting an automatic driving obstacle based on image recognition according to claim 4, wherein evaluating the reliability of a repair area in the global repair image, generating a repair reliability map, weighting image features according to the repair reliability map, and outputting an anti-repair artifact attention feature map, specifically comprising: Evaluating the content reconstruction reliability of each pixel position in the local repair image, and generating a repair reliability map corresponding to the global repair image space; Extracting a primary depth feature map from the global repair image through a feature extraction network, and upsampling the repair reliability map to the same spatial dimension as the primary depth feature map as a pixel-by-pixel attention weight; And multiplying each feature vector in the primary depth feature map with a corresponding attention weight to generate an anti-repair artifact attention feature map, wherein in the anti-repair artifact attention feature map, the feature amplitude of a position which is derived from a repair area and has low reliability is inhibited, and the feature amplitude of a clear area or a high-reliability repair area of an original image is maintained or enhanced.
  7. 7. The method for detecting an automated driving obstacle based on image recognition according to claim 6, wherein evaluating the content reconstruction reliability of each pixel position in the local repair image, generating a repair reliability map corresponding to the global repair image space, specifically comprises: And for each pixel position in the local restoration image, calculating a local structure similarity index corresponding to a corresponding pixel in the clear neighborhood image block, combining iteration convergence residual errors when the pixel position is reconstructed based on a physical degradation model inverse process, comprehensively calculating content reconstruction reliability, and generating a restoration reliability map consistent with the global restoration image spatial resolution.
  8. 8. The image recognition-based automatic driving obstacle detection method according to claim 1, wherein the obstacle detection is performed using a time-series cooperative attention mechanism based on the anti-repair artifact attention profile of the current frame and the history frame, and the output of the obstacle list specifically includes: stacking the anti-repair artifact attention feature map of the current frame and the historical anti-repair artifact attention feature map of the previous continuous N frames read from the buffer memory according to time sequence to form a feature map sequence, and inputting the feature map sequence into a time sequence coding network; The time sequence coding network processes the feature map sequence through a multi-head self-attention mechanism of a time dimension, and calculates the correlation weight of each frame feature and all frame features in the sequence, so that space-time features are extracted; and inputting the space-time characteristics into a detection head network, generating a boundary box, a category and a confidence coefficient, and outputting the obstacle list with the initial judgment information of the motion track by associating the detection result of the current frame with the track history predicted based on the space-time characteristics.
  9. 9. The method of claim 8, wherein the multi-headed self-attention mechanism is characterized in that the correlation weights enhance portions of the spatiotemporal features associated with entities that agree with spatial locations and appearance features within consecutive multiframes, and suppress portions associated with transient repair residual artifacts that occur in only a single frame or a small number of frames.
  10. 10. An image recognition-based automatic driving obstacle detection system, characterized by being configured to implement the image recognition-based automatic driving obstacle detection method according to any one of claims 1 to 9, comprising: The degradation region detection and segmentation module is used for acquiring an original image of a current frame acquired by the vehicle-mounted camera, identifying an instantaneous local degradation region caused by non-fixed attachments in the image, and generating an accurate mask of the degradation region based on a physical degradation model and image gradient analysis; The image restoration and fusion module is used for extracting a clear neighborhood image block from an original image by utilizing the accurate mask, reconstructing the content of a mask region based on the inverse process of the physical degradation model to generate a local restoration image, fusing the local restoration image with an undegraded region of the original image, and outputting a global restoration image; The repair reliability evaluation and feature optimization module is used for evaluating the content reconstruction reliability of each pixel in the local repair image to generate a repair reliability map, and weighting a primary depth feature map extracted from the global repair image by taking the repair reliability map as an attention weight to generate an anti-repair artifact attention feature map; The time sequence cooperative obstacle detection module is used for inputting the anti-repair artifact attention characteristic graphs of the current frame and the historical frame into a time sequence coding network, extracting space-time characteristics through a cooperative attention mechanism to strengthen continuous motion entity characteristics and inhibit interference characteristics, and executing detection and outputting an obstacle list with motion track initial judgment information based on the space-time characteristics.

Description

Automatic driving obstacle detection method and system based on image recognition Technical Field The invention relates to the technical field of vehicle image processing, in particular to an automatic driving obstacle detection method and system based on image recognition. Background The environment perception capability of an automatic driving system is highly dependent on a visual sensor, and especially in a complex road scene, accurate and stable detection of dynamic obstacles is a key premise of safety decision. The current image recognition method based on deep learning has good effect under the condition of regular illumination. However, in practical applications, the onboard camera lens often suffers from transient, localized physical contamination or shielding, such as splashed water droplets, mud spots, or localized dirt during travel. Such problems result in blurring, distortion or loss of information in local areas of the image, while the surrounding majority of the image area remains clear. The prior art has focused mainly on dealing with global image quality degradation such as low light enhancement, rain and fog removal, etc. For the problem of local physical degradation, the main scheme has obvious limitation of simply neglecting or globally filtering, namely, smoothing the local degradation as noise, so that the barrier information of the area is permanently lost, and the risk of missed detection is caused. The general image restoration is directly applied, although textures can be reconstructed, the consideration of a physical degradation model is lacked, structural artifacts exist in a restoration result, and the artifacts are easily misjudged as real obstacles by a subsequent detection network, so that false detection is caused. More importantly, the existing process generally regards repair and detection as two independent stages, the uncertainty of the result of the repair module is not evaluated, and the detection module lacks discrimination capability for potential errors introduced by repair, so that a perceived reliability vulnerability is formed. Therefore, a systematic solution that can specifically cope with instantaneous local physical occlusion, evaluate repair quality while repairing images, and enable detection of network adaptive trust reliable information is needed to fill the technical gap of current automatic driving visual perception in such cold but high-risk scenes. Disclosure of Invention The invention aims to provide an automatic driving obstacle detection method and system based on image recognition, which are used for solving the obstacle detection problem in the prior art caused by the degradation of instantaneous local images due to water drops, mud points and the like. In order to solve the technical problems, the invention specifically provides the following technical scheme: the automatic driving obstacle detection method based on image recognition comprises the following steps: S1, detecting an instantaneous local degradation area in an original image of a current frame, and generating an accurate mask; S2, reconstructing and fusing image content based on the accurate mask, and outputting a global repair image; S3, evaluating the reliability of the repair area in the global repair image, generating a repair reliability map, weighting image features according to the repair reliability map, and outputting an anti-repair artifact attention feature map; S4, based on the anti-repair artifact attention feature map of the current frame and the historical frame, using a timing coordinated attention mechanism to detect the obstacle, and outputting an obstacle list. As a preferred embodiment of the present invention, the S1 specifically includes: s11, reading and caching an original image of a current frame in real time through at least one optical camera interface in the vehicle-mounted looking-around system; s12, identifying an instantaneous local degradation area caused by non-fixed attachments in the original image of the current frame; S13, based on a point spread function physical degradation model of water drops and mud points, fuzzy kernel estimation is carried out on the identified candidate instantaneous local degradation areas, and the false detection areas are removed by combining gradient direction consistency verification in image gradient analysis, so that an accurate mask with accurate boundaries is generated. As a preferred embodiment of the present invention, the S12 specifically includes: comparing the pixel level difference of the original image of the current frame with that of a historical clear reference frame before a preset frame number, and identifying the instantaneous local degradation area which accords with the characteristic of non-fixed attachments by combining the change mode of the local gradient amplitude of the image; Wherein the change pattern comprises anisotropic attenuation characteristics of an edge diffusion gradient, and the