Search

CN-122024492-A - Full bridge deck traffic flow load space-time distribution monitoring method and system based on end-to-end

CN122024492ACN 122024492 ACN122024492 ACN 122024492ACN-122024492-A

Abstract

The invention provides an end-to-end-based full bridge deck traffic flow load space-time distribution monitoring method and system, which relate to the technical field of vision and laser fusion and comprise the following steps of acquiring a vehicle 2D image, 3D point cloud data, vehicle weight information of an upper bridge vehicle and corresponding moments; registering the vehicle 2D image and the 3D point cloud data under the same bridge floor coordinate system, constructing an end-to-end model according to the registered vehicle 2D image and the 3D point cloud data, inputting the vehicle 2D image acquired in real time into the end-to-end model to obtain a vehicle space position, forming a path sequence of the vehicle positions of the same vehicle (the same 2D image features or the same high-dimensional feature vector group in the 3D space constructed by the 2D image) to obtain a vehicle running track, and distributing vehicle weight information to the corresponding vehicle through time synchronization to obtain the vehicle load space-time distribution of the whole bridge floor. The method is easy to deploy, is easy to flexibly integrate in the bridge distributed monitoring system, and is easy to construct a distributed bridge group monitoring network.

Inventors

  • DAN DANHUI
  • Hu Jinnan

Assignees

  • 同济大学

Dates

Publication Date
20260512
Application Date
20260413

Claims (8)

  1. 1. The full bridge deck traffic flow load space-time distribution monitoring method based on end-to-end is characterized by comprising the following steps of: Acquiring a vehicle 2D image, 3D point cloud data and vehicle weight information of an upper bridge vehicle and corresponding moments, wherein the vehicle 2D image and the 3D point cloud data are provided with time tags; registering the vehicle 2D image and the 3D point cloud data under the same bridge floor coordinate system, and constructing an end-to-end model according to the registered vehicle 2D image and 3D point cloud data; inputting a vehicle 2D image acquired in real time into an end-to-end model to obtain a vehicle position; And forming a path sequence from the vehicle positions of the same vehicle to obtain a vehicle running track, distributing vehicle weight information to the corresponding vehicle through time synchronization, and obtaining vehicle load space-time distribution of the whole bridge deck, wherein the same vehicle has the same 2D image characteristics or the same high-dimensional characteristic vector group in a 3D space constructed by the 2D images.
  2. 2. The end-to-end based full bridge deck traffic flow load space-time distribution monitoring method of claim 1, wherein registering the vehicle 2D image and the 3D point cloud data under the same bridge deck coordinate system comprises the steps of: Performing instance segmentation on the 2D image to obtain a vehicle 2D mask; setting a coordinate limit value in a 3D point cloud space to obtain vehicle 3D point cloud data; And (3) taking the bridge floor world coordinate system as a reference, matching the matched vehicle 2D mask with the vehicle 3D point cloud data by using a time tag and a converted 3D space position of the vehicle 2D mask as well as a time tag and a 3D space position of the vehicle 3D point cloud data, and registering the matched vehicle 2D mask with the vehicle 3D point cloud data to obtain a training data set, wherein the training data set comprises a plurality of vehicle 2D masks and vehicle 3D point cloud data.
  3. 3. The end-to-end based full bridge deck traffic flow load space-time distribution monitoring method according to claim 2, wherein constructing an end-to-end model according to the registered vehicle 2D image and 3D point cloud data specifically comprises: Respectively mapping the mutually registered vehicle 2D mask and vehicle 3D point cloud data into two-dimensional multi-scale features and initial three-dimensional features, constructing a high-dimensional feature vector group by using a plurality of groups of two-dimensional multi-scale features, and lifting the extracted 2D image features to the representation of a 3D space to obtain mapped three-dimensional features; And performing supervision training on the mapped three-dimensional features and the initial three-dimensional features, and constructing an end-to-end model.
  4. 4. The end-to-end based full bridge deck traffic flow load space-time distribution monitoring method according to claim 1, wherein the vehicle positions of the same vehicles form a path sequence to obtain a vehicle running track, and the vehicle weight information is distributed to the corresponding vehicles through time synchronization to obtain the full bridge deck vehicle load space-time distribution, and the method specifically comprises: Acquiring a first motion track of a vehicle through real-time bridge deck image data of a first pair of camera vision ranges arranged at the bridge deck starting position, so as to acquire space-time distribution of vehicle loads in the first pair of camera vision ranges; Obtaining a second motion track of the vehicle through real-time bridge deck image data of a second pair of camera visual field ranges which are arranged along the bridge length and have an overlapping area with the first pair of cameras, so as to obtain space-time distribution of vehicle loads in the second pair of camera visual field ranges; Matching the vehicle 2D images of the overlapping areas of the first view field range and the second view field range, butting the first motion trail and the second motion trail with the same mapping three-dimensional characteristics, and so on, splicing the vehicle trail in the plurality of pairs of cameras arranged along the bridge length to form a complete vehicle bridge deck running trail, and obtaining the vehicle load space-time distribution of the whole bridge deck.
  5. 5. Full bridge floor traffic flow load space-time distribution monitoring system based on end to end, its characterized in that includes: The video monitoring unit is a plurality of cameras arranged along the bridge and is used for providing 2D image data covering the whole bridge deck, and overlapping areas exist in the visual fields of two adjacent cameras along the length direction of the bridge; The laser radar is temporarily installed at the same or similar positions of cameras arranged along the bridge length in the learning training period and used for providing 3D point cloud data matched with the 2D images shot by the video monitoring unit, and the laser radar is removed after the learning target is achieved; The road surface dynamic weighing unit is arranged on the bridge deck of the approach bridge close to the initial position of the target bridge and is used for acquiring the weight information of the on-bridge vehicle; The bridge deck traffic flow load AI perception large model system is used for monitoring and training of a vehicle end-to-end recognition positioning model based on pure vision, endowing vehicle load data measured by a road surface dynamic weighing unit with vehicle pixel targets at corresponding space-time positions through a precise space-time registration technology, and obtaining vehicle load space-time distribution of the whole bridge deck through vehicle target matching between adjacent cameras.
  6. 6. The end-to-end based full bridge deck traffic flow load space-time distribution monitoring system of claim 5, wherein the video monitoring units are symmetrically arranged above the bridge deck by L-shaped uprights or portal frames with longitudinal central axes of the bridge deck being symmetrical.
  7. 7. The end-to-end based full bridge deck traffic flow load space-time distribution monitoring system according to claim 5, wherein the laser radar sampling frequency is equal to or higher than 10Hz.
  8. 8. The end-to-end based full bridge deck traffic flow load space time distribution monitoring system of claim 5, wherein the road surface dynamic weighing unit measures the total weight, axle weight, speed and transit time of the vehicle in real time.

Description

Full bridge deck traffic flow load space-time distribution monitoring method and system based on end-to-end Technical Field The invention relates to the technical field of vision and laser fusion, in particular to an end-to-end-based full bridge deck traffic flow load space-time distribution monitoring method and system. Background In a bridge monitoring environment, the existing vehicle target recognition and positioning method is often difficult to directly obtain the spatial position of a vehicle in a bridge deck coordinate system through data acquired by a sensor. Specifically, three main stream vehicle target recognition and positioning methods are based on machine vision, laser radar and microwave radar, respectively, and all have the problem. Although the perception mechanisms of the three methods are different, a two-stage processing frame of 'target detection and coordinate conversion' is followed, namely, after the vehicle target detection is finished by the methods by utilizing respective vision acquisition and perception technologies, the coordinate conversion is realized by means of targets, bridge floor reference points or other references arranged on the bridge floor, and the vehicle position in the image coordinate system is converted to the bridge floor coordinate system. The coordinate conversion process depending on the external calibration object essentially introduces an intermediate conversion link for the vehicle positioning process, namely, firstly completing vehicle identification in a perception domain, then realizing coordinate mapping of a geometric domain through a target/reference object, and finally obtaining the space position of the vehicle under a bridge deck coordinate system. The intermediate link not only increases the extra workload of on-site layout, calibration and data processing, but also causes that the vehicle target recognition and positioning method cannot directly optimize the position error under the bridge deck coordinate system end to end through a unified loss function, and further the internal mapping relation between the image appearance and the real space position is difficult to fully learn. In addition, from the technical characteristics of each method, the method based on machine vision has the advantages of optimal economy but core defects of depth information deficiency, and the method based on laser radar and microwave radar can provide depth information but lacks of vehicle appearance and texture characteristics, so that the target identification robustness is insufficient, and the two methods are required to bear heavy data labeling work. It is worth noting that the 3D semantic occupation prediction method in the automatic driving field provides an important thought for reasoning the three-dimensional space position of the vehicle target from the two-dimensional image end to the end, but the design of the method is initially to achieve the purposes of complete environment perception, high-dimensional occupation modeling, multi-semantic class expression and the like, and the method has a significant difference from the task requirements of bridge structure traffic flow load monitoring and cannot be directly reused. Disclosure of Invention The invention aims to provide an end-to-end based full bridge deck traffic flow load space-time distribution monitoring method and system, which break through the technical bottleneck that the existing method relies on bridge deck layout targets, reference points and other references to realize coordinate conversion, realize vehicle end-to-end identification and space positioning by taking multi-view camera images as input, and simplify the monitoring flow of bridge traffic flow load space-time distribution in a random multi-traffic flow scene. In order to achieve the above purpose, the invention provides an end-to-end based full bridge deck traffic flow load space-time distribution monitoring method, which comprises the following steps: Acquiring a vehicle 2D image, 3D point cloud data and vehicle weight information of an upper bridge vehicle and corresponding moments, wherein the vehicle 2D image and the 3D point cloud data are provided with time tags; registering the vehicle 2D image and the 3D point cloud data under the same bridge floor coordinate system, and constructing an end-to-end model according to the registered vehicle 2D image and 3D point cloud data; inputting a vehicle 2D image acquired in real time into an end-to-end model to obtain a vehicle position; The vehicle positions of the same vehicles (the 2D image features are the same or the high-dimensional feature vector sets in the 3D space constructed by the 2D images are the same) form a path sequence to obtain vehicle running tracks, and vehicle weight information is distributed to the corresponding vehicles through time synchronization to obtain vehicle load space-time distribution of the whole bridge deck. Preferably, registering the vehicle 2