CN-121999315-A - Training method for abnormal event detection model, traffic event detection method and device, storage medium, program product and computer equipment

CN121999315ACN 121999315 ACN121999315 ACN 121999315ACN-121999315-A

Abstract

The application discloses a training method of an abnormal event detection model, a traffic event detection method and device, a storage medium, a program product and computer equipment, wherein the method comprises the steps of obtaining real data; the method comprises the steps of generating real data, generating synthetic data, calling a discriminator, generating a data true-false label set based on the data set to be discriminated, calling an event detection discriminator, obtaining an event label set based on the data set to be detected, wherein first data to be detected is constructed by historical frame data, second data to be detected is constructed by historical frame data and the synthetic data, determining a loss function of a model to be trained based on the data true-false label set, the event label set and the data type label set, carrying out iterative training on the model to be trained based on the loss function of the model to be trained, and enabling the event detection discriminator to be used as an abnormal event detection model after the iterative training is completed, so that detection and identification accuracy of the abnormal event detection model can be improved.

Inventors

SONG KUN
CHEN JINXUAN
LIU JINING
HUANG CHAOPING
HUANG XINGWEI
WU SHIFAN
WEN SHANGDONG
ZHENG WEINAN
XU HAO
LIN JIAO

Assignees

中国移动通信集团广东有限公司
中移湾区(广东)创新研究院有限公司
中国移动通信集团有限公司

Dates

Publication Date: 20260508
Application Date: 20260106

Claims (10)

1. The training method of the abnormal event detection model is characterized in that the model to be trained comprises a generator, a discriminator and an event detection discriminator, and comprises the following steps: acquiring real data, wherein the real data comprises historical frame data and a data type tag set corresponding to the historical frame data; invoking the generator to generate synthetic data based on the real data; invoking the discriminator, and generating a data authenticity label set based on a data set to be discriminated comprising the historical frame data and the synthesized data; Invoking the event detection discriminator to obtain an event tag set based on a to-be-detected data set containing first to-be-detected data and second to-be-detected data, wherein the first to-be-detected data is constructed by the historical frame data, and the second to-be-detected data is constructed by the historical frame data and the synthesized data; Determining a loss function of the model to be trained based on the data authenticity tag set, the event tag set and the data type tag set; And carrying out iterative training on the model to be trained based on the loss function of the model to be trained, wherein the event detection discriminator is used as the abnormal event detection model after the iterative training is completed.
2. The training method of claim 1, wherein, The generator includes a first timing encoder, a first type encoder, a first feature encoder, and a pre-header; The first timing encoder is used for converting timing data corresponding to the historical frame data into first timing characteristics; The first type encoder is configured to convert the set of data type tags into first type features; The first feature encoder is configured to generate a first depth feature according to the first timing feature and the first type feature; the prediction head is used for predicting according to the first depth characteristic so as to generate the synthesized data.
3. The training method of claim 1, wherein, The discriminator comprises a second time sequence encoder, a second type encoder, a second characteristic encoder and a first multi-layer perceptron MLP layer; The second time sequence encoder is used for converting time sequence data corresponding to the data set to be judged into second time sequence characteristics; the second type encoder is used for converting the type tag set of the data set to be discriminated corresponding to the data set to be discriminated into a second type characteristic; The second feature encoder is configured to generate a second depth feature according to the second timing feature and the second type feature; the first MLP layer is used for generating a discrimination probability according to the second depth characteristic, and the discrimination probability is used for constructing the data true-false label set.
4. The training method of claim 1, wherein, The event detection discriminator comprises a third time sequence encoder, a third feature encoder and a second MLP layer; The third time sequence encoder is used for converting time sequence data corresponding to the data set to be detected into a third time sequence characteristic; The third feature encoder is configured to generate a third depth feature according to the third timing feature; The second MLP layer is used for generating a prediction category corresponding to the data set to be detected by using an activation function according to the third depth feature, and the prediction category is used for constructing the event tag set.
5. The training method of claim 1, wherein the loss function of the model to be trained comprises the loss functions of the generator, the arbiter, and the event detection arbiter, respectively; The determining the loss function of the model to be trained based on the data authenticity tag set, the event tag set and the data type tag set comprises the following steps: Determining a loss function of the generator based on the data authenticity tab set, the event tab set and the data type tab set; determining a loss function of the discriminator based on the data authenticity tag set; And determining a loss function of the event detection discriminator based on the event tag set and the data type tag set.
6. A traffic event detection method, comprising: Acquiring traffic frame data to be detected; Inputting the traffic frame data to be detected into an abnormal event detection model to obtain an event tag output by the abnormal event detection model, wherein the abnormal event detection model is obtained according to the training method of any one of claims 1-5.
7. A traffic event detection device, comprising: the frame data acquisition module is configured to acquire traffic frame data to be detected; The detection module is configured to input the traffic frame data to be detected into an abnormal event detection model to obtain an event tag output by the abnormal event detection model, wherein the abnormal event detection model is obtained according to the training method of any one of claims 1-5.
8. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the method of any of claims 1-6.
9. A computer program product comprising computer instructions which, when executed by a processor, implement the method of any of claims 1-6.
10. A computer device comprising a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, the processor implementing the method of any of claims 1-6 when the computer program is executed.

Description

Training method for abnormal event detection model, traffic event detection method and device, storage medium, program product and computer equipment Technical Field The present application relates to the field of artificial intelligence, and in particular, to a training method for an abnormal event detection model, a traffic event detection method and apparatus, a storage medium, a program product, and a computer device. Background In the related art, artificial intelligence models typically require enough training samples to train. However, for some abnormal events, the abnormal events are of sporadic nature, and it is difficult to collect a sufficient amount of abnormal event samples in real life for training a model, so that the detection and identification accuracy of the trained model on the abnormal events is not high. Disclosure of Invention In order to solve the technical problems, the embodiment of the application provides a training method of an abnormal event detection model, a traffic event detection method and device, a storage medium, a program product and computer equipment, which can generate a large number of rich abnormal event training samples and improve the detection and identification accuracy of the abnormal event detection model. In a first aspect, an embodiment of the present application provides a training method for an abnormal event detection model, where a model to be trained includes a generator, a discriminator, and an event detection discriminator, and the training method includes: acquiring real data, wherein the real data comprises historical frame data and a data type tag set corresponding to the historical frame data; invoking the generator to generate synthetic data based on the real data; invoking the discriminator, and generating a data authenticity label set based on a data set to be discriminated comprising the historical frame data and the synthesized data; Invoking the event detection discriminator to obtain an event tag set based on a to-be-detected data set containing first to-be-detected data and second to-be-detected data, wherein the first to-be-detected data is constructed by the historical frame data, and the second to-be-detected data is constructed by the historical frame data and the synthesized data; Determining a loss function of the model to be trained based on the data authenticity tag set, the event tag set and the data type tag set; And carrying out iterative training on the model to be trained based on the loss function of the model to be trained, wherein the event detection discriminator is used as the abnormal event detection model after the iterative training is completed. Optionally, the generator includes a first timing encoder, a first type encoder, a first feature encoder, and a pre-header; The first timing encoder is used for converting timing data corresponding to the historical frame data into first timing characteristics; The first type encoder is configured to convert the set of data type tags into first type features; The first feature encoder is configured to generate a first depth feature according to the first timing feature and the first type feature; the prediction head is used for predicting according to the first depth characteristic so as to generate the synthesized data. Optionally, the arbiter comprises a second timing encoder, a second type encoder, a second feature encoder, and a first multi-layer perceptron MLP layer; The second time sequence encoder is used for converting time sequence data corresponding to the data set to be judged into second time sequence characteristics; the second type encoder is used for converting the type tag set of the data set to be discriminated corresponding to the data set to be discriminated into a second type characteristic; The second feature encoder is configured to generate a second depth feature according to the second timing feature and the second type feature; the first MLP layer is used for generating a discrimination probability according to the second depth characteristic, and the discrimination probability is used for constructing the data true-false label set. Optionally, the event detection arbiter comprises a third timing encoder, a third feature encoder, and a second MLP layer; The third time sequence encoder is used for converting time sequence data corresponding to the data set to be detected into a third time sequence characteristic; The third feature encoder is configured to generate a third depth feature according to the third timing feature; The second MLP layer is used for generating a prediction category corresponding to the data set to be detected by using an activation function according to the third depth feature, and the prediction category is used for constructing the event tag set. Optionally, the loss function of the model to be trained includes the loss functions of the generator, the discriminator and the event detection discriminator; The determining the los