Search

CN-121982929-A - Traffic accident monitoring system and method based on image recognition

CN121982929ACN 121982929 ACN121982929 ACN 121982929ACN-121982929-A

Abstract

The application provides a traffic accident monitoring system and a method based on image recognition, which are used for determining continuous track information of each monitoring target, extracting static attribute and dynamic behavior characteristic of each monitoring target, constructing a semantic interaction graph, generating probability distribution of the monitoring target on a predicted track in a future short-time domain through the semantic interaction graph and a pre-trained track prediction model, further determining collision probability of the monitoring target on physical collision in the future short-time domain, determining a motion risk index according to instantaneous speed and acceleration of the monitoring target on the dynamic behavior characteristic, fusing the collision probability with the motion risk index to obtain a comprehensive risk index, and judging that high probability collision risk exists and triggering early warning when the comprehensive risk index exceeds a preset risk threshold. By adopting the scheme of the application, the self-adaptive conflict risk monitoring can be carried out on the intersections based on the deep fusion traffic scene semantics.

Inventors

  • SHI DAN
  • Zhou Feiqin

Assignees

  • 南京交通职业技术学院

Dates

Publication Date
20260505
Application Date
20251230

Claims (10)

  1. 1. The intersection traffic collision risk early warning method is applied to a traffic accident monitoring system based on image recognition and is characterized by comprising the following steps of: acquiring an overhead video stream of a monitoring intersection, and determining continuous track information of each monitoring target, wherein the monitoring targets comprise tracked vehicle, pedestrian and non-motor vehicle targets; Extracting static attribute and dynamic behavior characteristics of each monitoring target; For a monitoring target pair entering a preset conflict area, constructing a semantic interaction diagram according to corresponding static attribute, dynamic behavior characteristics and road weight rule semantic information of the conflict area, and generating probability distribution of a predicted track of the monitoring target pair in a short time domain in the future through the semantic interaction diagram and a pre-trained track prediction model; Determining the collision probability of the monitoring target pair in the future short-time domain based on the probability distribution of the predicted track, determining a motion risk index according to the instantaneous speed and the acceleration of the monitoring target pair in the dynamic behavior characteristic, and fusing the collision probability and the motion risk index to obtain a comprehensive risk index; and when the comprehensive risk index exceeds a preset risk threshold, judging that the high probability collision risk exists and triggering early warning.
  2. 2. The method of claim 1, wherein extracting static attributes and dynamic behavior features of each monitored target specifically comprises: identifying and extracting static properties of each monitoring target from the overhead video stream; dynamic behavior features of the corresponding monitored targets are determined based on the respective continuous track information.
  3. 3. The method of claim 1, wherein for the pair of monitoring targets entering the preset conflict area, constructing the semantic interaction graph according to the corresponding static attribute, dynamic behavior feature and road right rule semantic information of the conflict area specifically comprises: Creating corresponding graph nodes for each monitoring target entering a preset conflict area, and taking static attribute and dynamic behavior characteristics of each monitoring target as feature vectors of the graph nodes corresponding to each monitoring target; establishing edges between nodes according to the space-time distance between the monitoring target pairs, and injecting traffic rule constraint for the whole graph structure by combining road weight rule semantic information of the conflict area; And outputting a semantic interaction graph fusing the feature vectors, interaction relations and rule semantics of each monitoring target.
  4. 4. The method of claim 1, wherein generating a probability distribution of monitoring targets to predicted trajectories in a short-term future domain from the semantic interaction map and a pre-trained trajectory prediction model comprises: Coding a node characteristic matrix, an adjacent matrix and a graph global attribute dictionary contained in the semantic interactive graph into an input tensor of a pre-trained track prediction model; inputting the encoded tensor into the pre-trained track prediction model, wherein the track prediction model is based on a graph neural network architecture, aggregating node characteristics, edge weights and global semantic information through a multi-layer message transmission mechanism, and outputting a position offset predicted value of each monitoring target at a future discrete time point; And superposing the position offset predicted value and the current coordinates of the monitoring targets, determining a plurality of possible tracks in a short time domain in the future, outputting corresponding probability values for each possible track, organizing all the possible tracks and the probability values thereof, and forming probability distribution of the independent predicted track of each monitoring target.
  5. 5. The method of claim 1, wherein determining a collision probability of a physical collision of a monitoring target pair in a future short-time domain based on a probability distribution of the predicted trajectory comprises: Expanding each possible track of the predicted track of each monitoring target in the pair of monitoring targets into a geometric area which changes with time in a future short-time domain, thereby converting the discrete track line into a continuous space-time probability volume; Calculating the intersection of the space-time probability volumes of the monitoring target pairs in the future, and if the intersection is not empty, identifying that the pair of monitoring targets has predicted track conflicts at the corresponding future time and space position; traversing all geometrical areas with intersections, weighting and calculating the total probability of the monitoring target on at least one physical conflict in the whole future short-time domain according to probability values of corresponding track modes causing the conflict, and outputting the total probability as the conflict probability.
  6. 6. The method of claim 1, wherein determining the risk of movement indicator based on the instantaneous speed and acceleration of the monitored object in relation to the dynamic behavior feature comprises: Determining the relative speed and the relative acceleration of the monitoring target pair at the current moment according to the instantaneous speed and the acceleration of the monitoring target pair at the current moment in the dynamic behavior characteristics; Acquiring current position coordinates of the monitoring target pair based on continuous track information, and further determining the relative position of the monitoring target pair; determining a conflict time index for quantifying risk urgency according to the relative position, the relative speed and the relative acceleration of the monitoring target pair; And carrying out standardization processing on the conflict time index, mapping the conflict time index into a scalar value related to risks, and taking the scalar value as a motion risk index representing the current moment and based on the motion state.
  7. 7. The method of claim 1, wherein fusing the collision probability and the athletic risk indicator to obtain a composite risk indicator comprises: carrying out fusion calculation on the collision probability and the motion risk index according to a predefined fusion function; And performing range constraint processing on the original value obtained by fusion calculation, and outputting a comprehensive risk index representing the overall collision risk of the monitoring target.
  8. 8. The system comprises an intersection traffic conflict risk early warning unit, and is characterized in that the intersection traffic conflict risk early warning unit comprises: The acquisition module is used for acquiring the overhead video stream of the monitoring intersection and determining continuous track information of each monitoring target, wherein the monitoring targets comprise tracked vehicles, pedestrians and non-motor vehicle targets; The processing module is used for extracting static attribute and dynamic behavior characteristics of each monitoring target; The processing module is further used for constructing a semantic interaction graph for a monitoring target pair entering a preset conflict area according to corresponding static attribute, dynamic behavior characteristics and road weight rule semantic information of the conflict area, and generating probability distribution of a predicted track of the monitoring target pair in a future short-time domain through the semantic interaction graph and a pre-trained track prediction model; The processing module is further used for determining the collision probability of the physical collision of the monitoring target in the future short-time domain based on the probability distribution of the predicted track, determining a motion risk index according to the instantaneous speed and the acceleration of the monitoring target in the dynamic behavior characteristic, and fusing the collision probability and the motion risk index to obtain a comprehensive risk index; And the execution module is used for judging that the high probability collision risk exists and triggering early warning when the comprehensive risk index exceeds a preset risk threshold value.
  9. 9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor, when executing the computer program, implements the steps of the intersection traffic collision risk warning method of any one of claims 1 to 7.
  10. 10. A computer-readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the steps of the intersection traffic collision risk warning method according to any one of claims 1 to 7.

Description

Traffic accident monitoring system and method based on image recognition Technical Field The application relates to the technical field of image recognition, in particular to a traffic accident monitoring system and method based on image recognition. Background Along with the rapid development of intelligent traffic, real-time monitoring of intersections by using a computer vision technology has become an important means for improving road traffic safety, and the traditional intersection safety monitoring method mainly relies on a video-based vehicle detection and tracking technology, and determines collision risks by calculating simple kinematic indexes such as space-time distances (such as Time To Collision (TTC)) among tracks. The prior art scheme mainly has the following limitations that firstly, perception and understanding are disjointed, most methods only pay attention to physical tracks of targets, lack of deep understanding of behavior intentions (such as straight going, turning and yielding) and traffic rules (such as road weight and signal lamp constraint) of traffic participants, one decelerating vehicle can be intented to yielding and can also be forced to stop due to congestion, potential risk differences of the vehicles can not be distinguished only by tracks, so that risk misjudgment is caused, secondly, the risk assessment dimension is single, the prior method only carries out collision probability prediction based on historical track extrapolation, ignores urgency of current instantaneous motion states, or carries out instant risk judgment only depending on current kinematic indexes, lacks predictability of future risk evolution, is difficult to comprehensively describe continuous risk spectrums from potential conflict to urgent danger, and thirdly, has insufficient adaptability, early warning thresholds of a plurality of systems are usually statically preset, cannot be dynamically adjusted according to specific geometric structures of intersections, real-time weather (such as rain and fog) and illumination conditions (such as backlight), so that the situation of collision is difficult to be suitable for the situation of the traffic conditions under different environments, and the situation of high warning rate of the traffic conditions is based on the depth of the intersection. Disclosure of Invention Based on the image recognition, the application provides a traffic accident monitoring system and a traffic accident monitoring method based on image recognition, which are used for carrying out self-adaptive conflict risk monitoring on intersections based on deep fusion traffic scene semantics. In a first aspect, the present application provides a method for early warning traffic collision risk at an intersection, which is applied to a traffic accident monitoring system based on image recognition, and the method comprises the following steps: acquiring an overhead video stream of a monitoring intersection, and determining continuous track information of each monitoring target, wherein the monitoring targets comprise tracked vehicle, pedestrian and non-motor vehicle targets; Extracting static attribute and dynamic behavior characteristics of each monitoring target; For a monitoring target pair entering a preset conflict area, constructing a semantic interaction diagram according to corresponding static attribute, dynamic behavior characteristics and road weight rule semantic information of the conflict area, and generating probability distribution of a predicted track of the monitoring target pair in a short time domain in the future through the semantic interaction diagram and a pre-trained track prediction model; Determining the collision probability of the monitoring target pair in the future short-time domain based on the probability distribution of the predicted track, determining a motion risk index according to the instantaneous speed and the acceleration of the monitoring target pair in the dynamic behavior characteristic, and fusing the collision probability and the motion risk index to obtain a comprehensive risk index; and when the comprehensive risk index exceeds a preset risk threshold, judging that the high probability collision risk exists and triggering early warning. In some embodiments, extracting the static attribute and the dynamic behavior feature of each monitoring target specifically includes: identifying and extracting static properties of each monitoring target from the overhead video stream; dynamic behavior features of the corresponding monitored targets are determined based on the respective continuous track information. In some embodiments, for a monitoring target pair entering a preset conflict area, constructing a semantic interaction graph according to corresponding static attribute, dynamic behavior characteristics and road right rule semantic information of the conflict area specifically includes: Creating corresponding graph nodes for each monitoring target