Traffic accident prediction systems and methods
Traffic accident prediction systems and methods are provided. The traffic accident prediction systems and methods include an accident prediction model that utilizes a spatiotemporal attention-based multi-graph convolution neural network to predict the number of traffic accidents in a predetermined region over a predetermined period of time in order to assist with the efficient dispatch of public safety resources to respond to traffic accidents.
1 . A traffic accident prediction system comprising:
a memory; and
a processor in communication with the memory, the processor configured to:
receive information related to traffic accidents in an area of interest;
extract a feature matrix of features of interest from the received traffic information;
construct a plurality of graphical representations from the feature matrix of features of interest;
perform a multi-layer multi-graph convolution with a layer-wise propagation rule;
apply a spatial attention mechanism formula, wherein the spatial attention mechanism formula adaptively captures dynamic correlations of spatial dimensions between nodes in the area of interest;
pass a feature matrix containing spatial features extracted through the multi-layer multi-graph convolution and application of the spatial attention mechanism through a recurrent neural network,
use calculations of recurrent units through the recurrent neural network to develop to context vector (“C”), wherein the context vector C stores spatiotemporal information of an encoder of an accident prediction model;
pass the context vector, C, to a decoder, wherein the decoder uses C as an initial hidden state to decode output sequences using a decoding formula;
apply a temporal attention mechanism to adaptively assign different weights of importance to different time periods;
create a feature matrix of temporal attention mechanism; and
use the feature matrix of temporal attention mechanism to determine a predicted number of traffic accidents in a given area over a predetermined amount of time.
2 . The traffic accident prediction system of claim 1 , wherein the processor is configured to preprocess the received information for use in the accident prediction model, by extracting a feature matrix of features of interest from the received information related to traffic accidents.
3 . The traffic accident prediction system of claim 1 , wherein the accident prediction model combines multiple graphical representations of the received information related to traffic accidents using a graph convolutional network.
4 . The traffic accident prediction system of claim 1 , wherein the accident prediction model performs a mapping function to predict an accident number count for a predetermined area over a predetermined period of time.
5 . The traffic accident prediction system of claim 1 , wherein the accident prediction model is an attention-based multi-graph convolutional network.
6 . The traffic accident prediction system of claim 1 , wherein the accident prediction model is comprised of a plurality of machine learning models.
7 . The traffic accident prediction system of claim 1 , wherein the accident prediction model outputs the predicted amount of traffic accidents in a predetermined geographic region over a predetermined period of time in the form of a traffic accident prediction map.
8 . A traffic accident prediction method comprising:
receiving information related to traffic accidents in an area of interest;
extracting a feature matrix of features of interest from the received traffic information;
constructing a plurality of graphical representations from the feature matrix of features of interest;
performing a multi-layer multi-graph convolution with a layer-wise propagation rule;
applying a spatial attention mechanism formula, wherein the spatial attention mechanism formula adaptively captures dynamic correlations of spatial dimensions between nodes in the area of interest;
passing a feature matrix containing spatial features extracted through the multi-layer multi-graph convolution and application of the spatial attention mechanism through a recurrent neural network,
using calculations of recurrent units through the recurrent neural network to develop to context vector (“C”), wherein the context vector C stores spatiotemporal information of an encoder of an accident prediction model;
passing the context vector, C, to a decoder, wherein the decoder uses C as an initial hidden state to decode output sequences using a decoding formula;
applying a temporal attention mechanism to adaptively assign different weights of importance to different time periods;
creating a feature matrix of temporal attention mechanism; and
using the feature matrix of temporal attention mechanism to determine a predicted number of traffic accidents in a given area over a predetermined amount of time.
9 . The traffic accident prediction method of claim 8 , wherein the method further comprises, extracting a feature matrix from the received information related to traffic accidents which is used to construct multiple graphical representations of the received information.
10 . The traffic accident prediction method of claim 8 , wherein receiving information related to traffic accidents, includes receiving a feature matrix of features of interest for a predetermined area and over a predetermined time.
11 . The traffic accident prediction method of claim 8 , wherein constructing multiple graphical representations of the received information includes an adaptive adjacency matrix, which self-learns hidden dependencies between entries in a feature matrix included in the received information.
12 . The traffic accident prediction method of claim 8 , wherein the method further comprises, visually projecting a predicted number of traffic accidents on a traffic accident prediction map.
13 . The traffic accident prediction method of claim 8 , wherein training a machine learning model comprises:
extracting features of interest from the received information;
performing a multi-graph, multi-layer convolution;
applying a spatial attention mechanism;
implementing a recurrent network to update future calculations based on past calculations;
supplying the machine learning model with known future information applying a temporal attention mechanism; and
outputting a predicted number of traffic accidents for a predetermined area over a predetermined period of time.
14 . The traffic accident prediction method of claim 13 , wherein implementing a recurrent network includes a long term short memory neural network.
15 . The traffic accident prediction method of claim 13 , wherein outputting a predicted number of traffic accidents for a predetermined area over a predetermined period of time includes visually projecting the predicted number of traffic accidents on a traffic accident prediction map.