CN-120913178-B - Automatic driving obstacle recognition method for road scene of complex scenic spot

CN120913178BCN 120913178 BCN120913178 BCN 120913178BCN-120913178-B

Abstract

The invention discloses an automatic driving obstacle recognition method for a complex scenic spot road scene, which comprises the steps of collecting an original image of a scenic spot road, inputting a multi-scale feature extraction network to generate a multi-scale feature image, inputting the multi-scale feature image into a multi-branch decoder to respectively obtain an obstacle region segmentation result, an obstacle position boundary frame coordinate and an obstacle type classification probability, combining surrounding environment information of road geometric forms, weather and signal lamp states, weighting the features by adopting an attention mechanism to generate an initial obstacle significance image, detecting and processing a spatial overlapping recognition result of the image according to a preset obstacle type priority level to obtain an obstacle recognition result after conflict resolution, performing spatial consistency correction on the obstacle recognition result and the road path segmentation image, removing inter-frame jump based on time sequence smoothing, and finally outputting the recognition result. The invention effectively solves the problem of high-precision and real-time obstacle recognition of road scenes facing complex scenic spots.

Inventors

DU HUAN
Wei Yinxi
WANG SHUXIANG
ZHU YANAN

Assignees

南京万星汇智能科技有限公司

Dates

Publication Date: 20260508
Application Date: 20250801

Claims (7)

1. An automatic driving obstacle recognition method for a complex scenic spot road scene is characterized by comprising the following steps of: Collecting an original image of a scenic spot road, and performing noise suppression, brightness equalization and contrast enhancement to obtain an enhanced image with uniform resolution; inputting the enhanced image into a multi-scale feature extraction network, respectively extracting texture features, shape features and semantic features, and fusing the texture features, the shape features and the semantic features through a feature pyramid network to generate a multi-dimensional feature map; Inputting the multi-dimensional feature map into a multi-branch decoder to respectively obtain an obstacle region segmentation result, obstacle position boundary frame coordinates and obstacle type classification probability; Combining the road geometry, weather and surrounding environment information of signal lamp states, carrying out feature weighting on the barrier region segmentation result, the barrier position boundary frame coordinates and the barrier type classification probability by adopting an attention mechanism, and generating a barrier significance feature map; Detecting and processing a spatial overlapping recognition result in the obstacle significance characteristic diagram according to a preset obstacle type priority, reserving a high-priority target, and inhibiting a low-priority target to obtain a conflict resolved obstacle recognition result; Performing spatial consistency correction on the obstacle recognition result and the road path segmentation map, eliminating inter-frame jump based on time sequence smoothing, and finally outputting recognition results comprising the type, the position, the segmentation mask and the confidence coefficient of the obstacle; The method comprises the steps of obtaining a barrier salient feature map, namely, extracting surrounding environment information of scenic spot roads, including road geometry information, surrounding road landscape layout information, signal condition information and weather condition information, constructing a multi-mode fusion frame, converting barrier region segmentation results into space salient features, converting barrier position positioning results into position salient features, and converting barrier type classification results into semantic salient features; The obstacle type priority level strategy is utilized to correct an obstacle position locating result and an obstacle type classifying result in the obstacle saliency characteristic map, and the method comprises the steps of establishing an obstacle type priority level of a road scene for an obstacle type, wherein a tourist pedestrian is set to be a first priority, a traffic vehicle is set to be a second priority, a park maintenance device is set to be a third priority, and a fixed landscape facility is set to be a fourth priority; conflict detection is carried out on a plurality of obstacle recognition results with spatial overlapping obstacle significance characteristics, and when a plurality of obstacles of different types appear in the same space, result screening is carried out according to a priority level strategy; performing transcription processing on the recognition result of the obstacle of the fourth priority through a non-maximum transcription factor, and reserving the position positioning result and the type classification result of the obstacle of the first priority to obtain the recognition result of the obstacle after conflict resolution; the method comprises the steps of combining scenic spot road path segmentation information, carrying out consistency check on a corrected recognition result, and finally outputting scenic spot obstacle recognition results, wherein the scenic spot road is subjected to road path region segmentation through a segmentation network, a passable road region, a tourist walking region, a greening landscape region and a building region are generated, a road path segmentation map is generated, the corrected obstacle recognition result is subjected to space consistency check with the road path segmentation map, whether the recognition position of an obstacle is matched with a road function region which should appear or not is checked, the recognition result which does not meet the space consistency check is checked, inter-frame jumps are eliminated based on time sequence smoothing, and finally a complete obstacle recognition result comprising an obstacle type label, a position coordinate, a region segmentation mask and a confidence score is output.
2. The method for identifying an automated driving obstacle for a complex scenic spot roadway scene according to claim 1, wherein said enhanced roadway image comprises: performing bilateral filtering on the scenic spot road original image to suppress random noise while preserving edge details; adopting a self-adaptive histogram equalization method to carry out local contrast improvement on the area with uneven brightness distribution; And dynamically adjusting the gamma coefficient according to the global brightness statistics to enable the overall brightness of the enhanced image to tend to be balanced.
3. The method for identifying the automatic driving obstacle for the complex scenic spot road scene according to claim 1 or 2, wherein the acquisition of the scenic spot road original image meets the following environmental requirements, comprising: Respectively acquiring scenic spot road images under the conditions of sunny days, cloudy days, rainy days, foggy days and night; The collection batch under each meteorological condition at least comprises three illumination directions of opposite sunlight, opposite sunlight and side light; the scenic spot road image at least comprises a scenic spot inner trunk road image, a scenic spot inner branch road image, a scenic spot parking lot and scenic spot road images under different air conditions, wherein: The scenic spot road images under different meteorological conditions comprise, but are not limited to, strong light images under sunny conditions, weak light images under cloudy conditions, light reflection images under rainy conditions, blurred images under foggy conditions, obvious images under daytime illumination conditions, light supplementing images under dusk light changes, and low-light images under night artificial conditions.
4. The method for automatically driving obstacle recognition for complex scenic spot road scenes according to claim 1, wherein generating the multi-dimensional feature map comprises: taking the enhanced road image as input data, and extracting the layer-by-layer characteristics of the input image through a shallow layer image layer group, a middle layer image layer group and a deep layer image layer group in a multi-dimensional characteristic extraction network; the shallow layer image group extracts image edge texture features, the middle layer image group extracts shape structure features of the image, and the deep layer image group extracts semantic content features of the image; and fusing the features of different layer groups through a feature pyramid network, unifying the sizes of the feature graphs through up-sampling and down-sampling operation, and finally outputting a multi-dimensional feature graph containing multi-dimensional feature information.
5. The method for recognizing an automatic driving obstacle for a complex scenic spot road scene according to claim 4, wherein the multi-dimensional feature map is input to a multi-branch decoder, and an obstacle region segmentation result, an obstacle position location result and an obstacle type classification result are obtained through three decoding branches, respectively, wherein: The first decoding branch is an obstacle region segmentation branch, analyzes the image resolution on the multi-dimensional feature map through gradient inverse matrix operation, judges the attribution of an obstacle region for each pixel point through a pixel-level classifier, and generates a binarized obstacle region segmentation mask; The second decoding branch is an obstacle position positioning branch, converts the multi-dimensional feature map into a feature with a fixed length through global average pooling operation for analysis, and predicts the boundary frame coordinate information of the obstacle through multi-layer connection network regression; the third decoding branch is an obstacle classification type branch, global signal characteristics are extracted through characteristic aggregation operation, and probability distribution of each obstacle is output through a classifier network.
6. The method for recognizing the automatic driving obstacle for the complex scenic spot road scene according to claim 5, wherein the obstacle region segmentation result comprises describing obstacle region boundary information with pixel level precision, wherein each pixel point corresponds to a binarization tag for indicating whether the pixel point belongs to an obstacle region or not, and forming a complete obstacle segmentation mask diagram; The obstacle position locating result comprises obstacle space position information expressed in the form of boundary boxes, wherein each boundary box is determined by an upper left corner coordinate and a lower right corner coordinate, the accurate position range of the obstacle in an image is calibrated, and meanwhile, the confidence score of the boundary box is included and used for expressing the reliability degree of the locating result; The obstacle type classification result comprises classification labels of common obstacle types in scenic spot roads, wherein the classification labels at least comprise tourist pedestrian categories, traffic vehicle categories, park maintenance equipment categories and fixed landscape facility categories, and each category label corresponds to a probability value and is used for representing the possibility of the category to which the obstacle belongs.
7. The method for identifying an automatic driving obstacle for a complex scenic road scene according to claim 1, wherein the time-series smoothing process specifically comprises: Estimating the position and the speed of the obstacle of the adjacent frame based on a Kalman filtering algorithm; feeding back the estimation result to the current frame identification result, weakening instantaneous positioning jitter and updating confidence coefficient; and outputting a final obstacle recognition result after the time smoothing.

Description

Automatic driving obstacle recognition method for road scene of complex scenic spot Technical Field The invention relates to the technical field of computer vision and scenic spot intelligent traffic, in particular to an automatic driving obstacle recognition method for a road scene of a complex scenic spot. Background With the rapid development of artificial intelligence and computer vision technologies, the environment perception and obstacle recognition technology in an automatic driving system has become the core research direction of the intelligent traffic field, the traditional obstacle recognition method mainly relies on single sensor data processing, acquires environment information through radar, laser radar or cameras and other equipment, and adopts a basic image processing algorithm to carry out target detection and classification, however, with the rise of deep learning technology, a vision perception method based on a convolutional neural network gradually becomes the mainstream, particularly shows remarkable advantages in multi-target recognition tasks in complex road environments, at present, an obstacle recognition framework usually adopts an end-to-end deep learning architecture, combines a plurality of subtasks such as feature extraction, target detection and semantic segmentation, realizes accurate recognition of various targets such as vehicles, pedestrians and traffic signs in road scenes, and simultaneously, the application of the multi-mode fusion technology further improves the robustness of the system under severe weather conditions, and the introduction of time sequence information effectively relieves the problem of false recognition in single-frame detection. CN115230694a discloses a method for identifying an obstacle of an automatic driving vehicle, which mainly aims at structural road design, has insufficient adaptability in terms of complex topography, diversified obstacle types and dynamic environment changes of scenic roads, and is particularly easy to generate erroneous judgment when processing non-standard road geometric forms, although the obstacle screening is realized through Frenet coordinate system conversion. CN118587683a discloses an obstacle recognition method based on a generated countermeasure network, which improves model generalization capability through data enhancement, but the characteristic extraction process lacks specific consideration of special environmental factors of scenic spot roads, such as complex illumination conditions, changeable weather states and specific moving obstacles (tourists, sightseeing vehicles and the like) of scenic spots, so that the problems of insufficient recognition precision and poor real-time performance exist in practical application. In addition, the prior art generally lacks an effective fusion mechanism for multi-scale features, is difficult to simultaneously consider the accurate detection of a long-distance small target and the complete segmentation of a short-distance large target, and is lack of an effective conflict resolution strategy when processing space overlapping targets, and false detection and omission are easy to generate, so that the existing automatic driving obstacle recognition technology has the problems of insufficient adaptability to the road environment of a complex scenic spot, limited multi-scale feature fusion capability, lack of environmental context information utilization and imperfect processing mechanism of the space overlapping targets, and mainly solves the problems of high-precision and real-time obstacle recognition for the road scene of the complex scenic spot. Disclosure of Invention This section is intended to summarize some aspects of embodiments of the application and to briefly introduce some preferred embodiments, which may be simplified or omitted in this section, as well as the description abstract and the title of the application, to avoid obscuring the objects of this section, description abstract and the title of the application, which is not intended to limit the scope of this application. The present invention has been made in view of the above-described problems occurring in the prior art. In order to solve the technical problems, the invention provides the following technical proposal that the original image of the scenic spot road is collected and noise suppression, brightness equalization and contrast enhancement are carried out to obtain an enhanced image with uniform resolution; inputting the enhanced image into a multi-scale feature extraction network, respectively extracting texture features, shape features and semantic features, and fusing the texture features, the shape features and the semantic features through a feature pyramid network to generate a multi-scale feature map; inputting the multi-scale feature map into a multi-branch decoder to respectively obtain an obstacle region segmentation result, obstacle position boundary frame coordinates and obstacle typ