A traffic accident recognition method and system based on image processing

Abstract

The embodiment of the present application provides a traffic accident identification method and system based on image processing. Based on the feature distance between the traffic accident description features of the traffic collection image stream data and the traffic accident description features of the sample traffic accident data, one or more second-category traffic accident data corresponding to the traffic collection image stream data are determined. Compared with manually calibrating the second-category traffic accident data corresponding to the traffic collection image stream data, the collection speed of the second-category traffic accident data is improved, thereby contributing more second-category traffic accident data to the knowledge learning of the traffic accident identification network, so that the traffic accident identification network can effectively learn the difference between the traffic collection image stream data and the second-category traffic accident data, improve the observation performance of the traffic accident identification network for the noisy second-category traffic accident data, and then improve the traffic accident identification accuracy of the target traffic accident identification network.

Description

Technical Field

[0001] The present application relates to the technical field of traffic accident identification, and in particular to a traffic accident identification method and system based on image processing. Background Art

[0002] At present, we can combine video image intelligent analysis to realize the intelligent monitoring function of various traffic accidents, effectively promote the construction of smart traffic monitoring, and improve the traditional non-site illegal evidence collection of high labor costs and low illegal investigation rate by intelligently monitoring various traffic accident behaviors and uploading them in real time, thus enhancing the credibility of non-site law enforcement. Among the related technologies, how to improve the accuracy of traffic accident recognition based on image processing is a technical problem that needs to be solved urgently. Summary of the invention

[0003] In view of this, the purpose of the present application is to provide a traffic accident identification method and system based on image processing.

[0004] According to a first aspect of the present application, there is provided a traffic accident recognition method based on image processing, which is applied to a traffic accident recognition system based on image processing and is executed by a traffic accident recognition network, wherein the traffic accident recognition network includes a first traffic accident description unit and a second traffic accident description unit, and the method includes: Acquire traffic collection image stream data, and one or more first-category traffic accident data corresponding to the traffic collection image stream data, wherein the first-category traffic accident data represents traffic accident data corresponding to the traffic collection image stream data; Extracting traffic accident description features of the traffic collection image stream data according to the first traffic accident description unit, the traffic accident description features including trajectory feature vectors of dynamic moving elements and trajectory feature vectors of accident-related elements in the traffic collection image stream data, the trajectory feature vectors of the dynamic moving elements and the trajectory feature vectors of the accident-related elements at least including position information of the dynamic moving elements and the accident-related elements; Based on the traffic accident description features of the traffic collection image stream data and the traffic accident description features corresponding to each sample traffic accident data, x traffic accident data associated with the traffic collection image stream data are determined from each sample traffic accident data, the traffic accident description features corresponding to the sample traffic accident data are obtained based on the second traffic accident description unit, and x is a positive integer; From the x traffic accident data, determine one or more traffic accident data to be output as second-category traffic accident data corresponding to the traffic collection image stream data, wherein the second-category traffic accident data represents traffic accident data that does not correspond to the traffic collection image stream data; The traffic accident recognition network is trained based on the first type of traffic accident data and the second type of traffic accident data corresponding to the traffic collection image stream data to generate a target traffic accident recognition network, wherein the target traffic accident recognition network is used to recognize traffic accident data existing in any collected traffic collection image stream data.

[0005] In a possible implementation manner of the first aspect, the determining, from the x traffic accident data, one or more traffic accident data output as the second type of traffic accident data corresponding to the traffic collection image stream data includes: From the traffic accident data included in the x traffic accident data except the one or more first-category traffic accident data, one or more traffic accident data are determined to be output as second-category traffic accident data corresponding to the traffic collection image stream data.

[0006] In a possible implementation manner of the first aspect, determining, from the traffic accident data other than the one or more first-category traffic accident data included in the x traffic accident data, outputting one or more traffic accident data as the second-category traffic accident data corresponding to the traffic collection image stream data comprises: removing the one or more first-category traffic accident data from the x traffic accident data to generate one or more remaining traffic accident data; Based on the characteristic distances between the traffic collection image stream data and each of the remaining traffic accident data, determining one or more traffic accident data having the largest characteristic distances to the traffic collection image stream data; Outputting the one or more traffic accident data with the largest characteristic distance as the second type of traffic accident data corresponding to the traffic collection image stream data; Alternatively, based on the characteristic distances between the traffic collection image stream data and each of the remaining traffic accident data, y traffic accident data having the largest characteristic distances with the traffic collection image stream data are determined, where y is a positive integer less than x; The one or more first-category traffic accident data are removed from the y traffic accident data, and the remaining traffic accident data are output as second-category traffic accident data corresponding to the traffic collection image stream data.

[0007] 4. The method for traffic accident identification based on image processing according to claim 1, characterized in that the traffic accident identification network is trained based on the first type of traffic accident data and the second type of traffic accident data corresponding to the traffic collection image stream data to generate a target traffic accident identification network, comprising: Based on the first type of traffic accident data and the second type of traffic accident data corresponding to the traffic collection image stream data, determining a network learning error value of the traffic accident recognition network, wherein the network learning error value reflects the effectiveness of the traffic accident recognition network in identifying whether the traffic accident data is the first type of traffic accident data or the second type of traffic accident data; Based on the network learning error value, the first traffic accident description unit is trained to generate the target traffic accident recognition network, and the weight information of the second traffic accident description unit is locked.

[0008] In a possible implementation of the first aspect, determining the network learning error value of the traffic accident recognition network based on the first type of traffic accident data and the second type of traffic accident data corresponding to the traffic collection image stream data includes: For target traffic accident data, determining a network learning error value of the traffic accident recognition network based on a feature distance between traffic accident description features of the traffic collection image stream data and traffic accident description features of the target traffic accident data, and traffic accident type information of the target traffic accident data; Among them, the target traffic accident data is any one of the first type of traffic accident data or any one of the second type of traffic accident data corresponding to the traffic collection image stream data, and the traffic accident type information reflects that the target traffic accident data belongs to the first type of traffic accident data or the second type of traffic accident data.

[0009] In a possible implementation manner of the first aspect, after the first traffic accident description unit is trained based on the network learning error value to generate the target traffic accident identification network, the method further includes: If the traffic accident recognition network does not meet the network convergence requirement, the target traffic accident recognition network is continuously iterated and updated, and the process returns to the step of obtaining the traffic collection image stream data and continues the iteration; For the same traffic collection image stream data, the traffic accident description features of the traffic collection image stream data are different in different batches of network learning processes, and the second type of traffic accident data corresponding to the traffic collection image stream data are different in different batches of network learning processes.

[0010] In a possible implementation of the first aspect, based on the traffic accident description features of the traffic collection image stream data and the traffic accident description features corresponding to each sample traffic accident data, determining x traffic accident data associated with the traffic collection image stream data from each sample traffic accident data includes: Generate a traffic accident reference sequence based on the traffic accident description features of the sample traffic accident data in the upper layer of the traffic accident identification bitmap, wherein the traffic accident identification bitmap is used to record the traffic accident description features corresponding to each sample traffic accident data based on the feature distance between the multiple sample traffic accident data, and the traffic accident reference sequence is used to dynamically record the identification bit information of the sample traffic accident data associated with the traffic collection image stream data in the traffic accident identification bitmap during the determination process of the traffic accident data; For the zth identification bit information included in the traffic accident reference sequence, one or more neighbor identification bit information associated with the zth identification bit information in the traffic accident identification bitmap is added to the traffic accident reference sequence, the traffic accident description feature of the sample traffic accident data corresponding to the neighbor identification bit information is adjacent to the traffic accident description feature of the sample traffic accident data corresponding to the zth identification bit information, and z is a positive integer; Determine x pieces of associated identification information from the traffic accident reference sequence to generate an adjusted traffic accident reference sequence; The feature distance between the traffic accident description feature of the sample traffic accident data corresponding to the associated identification information and the traffic accident description feature of the traffic collected image stream data is smaller than the feature distance between the traffic accident description feature of the sample traffic accident data corresponding to other identification information not selected in the traffic accident reference sequence and the traffic accident description feature of the traffic collected image stream data; Starting again from the step of adding one or more neighbor identification bit information associated with the z-th identification bit information in the traffic accident identification bitmap into the traffic accident reference sequence, executing a subsequent determination process; If the adjusted traffic accident reference sequence is the same as the adjusted traffic accident reference sequence obtained by the subsequent determination process, the sample traffic accident data corresponding to the x identification bit information included in the adjusted traffic accident reference sequence are used as the x traffic accident data.

[0011] In a possible implementation of the first aspect, the method further includes: For any one of the sample traffic accident data, output the sample traffic accident data as pending traffic accident data; determining, according to the second traffic accident description unit, a traffic accident description feature of the pending traffic accident data; Based on the traffic accident description features of the pending traffic accident data and the traffic accident description features of the traffic accident data recorded in the traffic accident identification bitmap, determining the first identification bit information of the traffic accident description features of the pending traffic accident data in the traffic accident identification bitmap, and the recorded traffic accident data belongs to the sample traffic accident data; Determine k recorded traffic accident data with the smallest characteristic distance from the pending traffic accident data, and use the identification information corresponding to the k recorded traffic accident data as neighbor identification information associated with the first identification information, where k is a positive integer.

[0012] In a possible implementation manner of the first aspect, the acquiring of traffic collection image stream data and one or more first-category traffic accident data corresponding to the traffic collection image stream data includes: Acquire traffic collection image stream data and a plurality of prior-labeled traffic accident data corresponding to the traffic collection image stream data; One or more a priori labeled traffic accident data with a traffic accident degree not less than a first threshold value are extracted from the multiple a priori labeled traffic accident data, and determined as one or more first-category traffic accident data corresponding to the traffic collection image stream data.

[0013] According to the second aspect of the present application, a traffic accident identification system based on image processing is provided, wherein the traffic accident identification system based on image processing comprises a machine-readable storage medium and a processor, wherein the machine-readable storage medium stores machine-executable instructions, and when the processor executes the machine-executable instructions, the traffic accident identification system based on image processing implements the aforementioned traffic accident identification method based on image processing.

[0014] According to a third aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed, the aforementioned image processing-based traffic accident identification method is implemented.

[0015] Based on any of the above aspects, in the present application, in the process of knowledge learning of the traffic accident recognition network, one or more second-category traffic accident data corresponding to the traffic collection image stream data are determined based on the feature distance between the traffic accident description features of the traffic collection image stream data and the traffic accident description features of the sample traffic accident data. Compared with manually calibrating the second-category traffic accident data corresponding to the traffic collection image stream data, the collection speed of the second-category traffic accident data is improved, thereby contributing more second-category traffic accident data to the knowledge learning of the traffic accident recognition network, so that the traffic accident recognition network can effectively learn the difference between the traffic collection image stream data and the second-category traffic accident data, improve the observation performance of the traffic accident recognition network for the second-category traffic accident data, and thereby improve the traffic accident recognition accuracy of the target traffic accident recognition network. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 A flow chart of a traffic accident recognition method based on image processing provided in an embodiment of the present application; Figure 2 A schematic diagram of the component structure of a traffic accident identification system based on image processing for implementing the above-mentioned traffic accident identification method based on image processing provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme and advantages of the embodiment of the present application clearer, the technical scheme in the embodiment of the present application will be clearly and completely described below according to the drawings in the embodiment of the present application. It should be understood that the drawings in the present application only serve the purpose of explanation and description, and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn in real proportion. The flowchart used in this application shows the operations implemented according to some embodiments of the embodiment of the present application. It should be understood that the operations of the flowchart can be implemented out of sequence, and the steps without logical context can be reversed in order or implemented simultaneously. In addition, those skilled in the art can add one or more other operations to the flowchart under the guidance of the content of the present application, or destroy one or more operations from the flowchart.

[0019] In addition, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. According to the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0020] Figure 1 The flowchart of the traffic accident identification method based on image processing provided by the embodiment of the present application is shown. It should be understood that in other embodiments, the order of some steps in the traffic accident identification method based on image processing of the present embodiment can be shared with each other according to actual needs, or some steps can be omitted or maintained. The details of the traffic accident identification method based on image processing include: Step S110, acquiring traffic collection image stream data and one or more first-category traffic accident data corresponding to the traffic collection image stream data, wherein the first-category traffic accident data represents traffic accident data corresponding to the traffic collection image stream data.

[0021] In an alternative implementation, traffic collection image stream data may be used to represent image stream data to be trained for traffic accident recognition.

[0022] In an alternative implementation, a characteristic distance between the first type of traffic accident data corresponding to the traffic collection image stream data and the traffic collection image stream data is small.

[0023] In an alternative implementation, in the knowledge learning of the traffic accident recognition network, sample learning data is constructed through traffic collection image stream data and traffic accident data, that is, the sample learning data includes traffic collection image stream data and traffic accident data corresponding to the traffic collection image stream data.

[0024] In an alternative implementation, the sample learning data is divided into two categories: positive sample learning data and negative sample learning data. The positive sample learning data includes traffic collection image stream data and the first category of traffic accident data corresponding to the traffic collection image stream data. The negative sample learning data includes traffic collection image stream data and the second category of traffic accident data corresponding to the traffic collection image stream data. The second category of traffic accident data represents traffic accident data that does not correspond to the traffic collection image stream data. For the implementation of the second category of traffic accident data, see the following content.

[0025] For a certain traffic collection image stream data, the first type of traffic accident data corresponding to the traffic collection image stream data is compared with the second type of traffic accident data corresponding to the traffic collection image stream data, and the characteristic distance between the first type of traffic accident data and the traffic collection image stream data is higher than the characteristic distance between the second type of traffic accident data and the traffic collection image stream data. That is, the correlation between the first type of traffic accident data and the traffic collection image stream data is better, and the correlation between the second type of traffic accident data and the traffic collection image stream data is poor, or the second type of traffic accident data does not match the traffic collection image stream data.

[0026] In an alternative implementation, the first type of traffic accident data corresponding to the traffic collection image stream data can be pre-formed before the knowledge learning of the traffic accident recognition network begins, and the second type of traffic accident data corresponding to the traffic collection image stream data can be dynamically generated during the knowledge learning of the traffic accident recognition network. In an alternative implementation, in different knowledge learning stages, the second type of traffic accident data corresponding to the same traffic collection image stream data are not completely the same.

[0027] Next, the method for determining the first type of traffic accident data is described.

[0028] In an alternative implementation, the traffic collection image stream data used for knowledge learning of the traffic accident recognition network is past traffic collection image stream data, that is, traffic collection image stream data collected in the past period of time.

[0029] In an alternative implementation, one or more prior annotated traffic accident data corresponding to the past traffic collection image stream data can be determined based on the past traffic collection image stream data, thereby determining the traffic collection image stream data used for knowledge learning when outputting the one or more past traffic collection image stream data, and the one or more first-category traffic accident data corresponding to the traffic collection image stream data. The degree of the traffic accident of the prior annotated traffic accident data can be determined.

[0030] In an alternative implementation, the traffic collection image stream data corresponds to a plurality of first-category traffic accident data, and each first-category traffic accident data and the traffic collection image stream data can constitute a positive sample learning data.

[0031] In an alternative embodiment, the first traffic accident description unit in the traffic accident recognition network is used to describe the traffic accident on the traffic collection image stream data, and the second traffic accident description unit is used to describe the traffic accident on the traffic accident data. In an alternative embodiment, the network structures of the first traffic accident description unit and the second traffic accident description unit are similar, and the network parameters of the first traffic accident description unit and the second traffic accident description unit are different.

[0032] Step S120: extracting traffic accident description features from the traffic collection image stream data according to the first traffic accident description unit.

[0033] The knowledge learning in the traffic accident recognition network includes one or more knowledge learning stages. For any of the one or more knowledge learning stages, M traffic collection image stream data and one or more first-category traffic accident data corresponding to the M traffic collection image stream data can be obtained, and then the traffic accident description features corresponding to the M traffic collection image stream data are calculated respectively, so as to find the second-category traffic accident data corresponding to the M traffic collection image stream data.

[0034] In an alternative implementation, for example, M=3, in the xth knowledge learning stage, traffic collection image stream data 1, traffic collection image stream data 2, and traffic collection image stream data 3 are selected; in the x+1th knowledge learning stage, traffic collection image stream data 1, traffic collection image stream data 5, and traffic collection image stream data 6 are selected. For example, in each knowledge learning stage, M traffic collection image stream data are randomly selected from a plurality of traffic collection image stream data, and the M traffic collection image stream data and the traffic accident data (the traffic accident data includes the first type of traffic accident data and the second type of traffic accident data) corresponding to each traffic collection image stream data are used as sample learning data used in the knowledge learning stage.

[0035] Step S130, based on the traffic accident description features of the traffic collection image stream data and the traffic accident description features corresponding to each sample traffic accident data, x traffic accident data associated with the traffic collection image stream data are determined from each sample traffic accident data, the traffic accident description features corresponding to the sample traffic accident data are obtained based on the second traffic accident description unit, and x is a positive integer.

[0036] In an alternative embodiment, the traffic accident description features corresponding to the sample traffic accident data are called traffic accident description features of the sample traffic accident data, and the traffic accident description features include trajectory feature vectors of dynamic moving elements and trajectory feature vectors of accident-related elements in the traffic collection image stream data, and the trajectory feature vectors of the dynamic moving elements and the trajectory feature vectors of the accident-related elements include at least the position information of the dynamic moving elements and the accident-related elements.

[0037] In an alternative implementation, the traffic accident description features corresponding to one or more sample traffic accident data may also be generated in the knowledge learning of the traffic accident identification network. For example, in the first knowledge learning stage of the knowledge learning of the traffic accident identification network, the traffic accident description features corresponding to one or more sample traffic accident data are determined by the second traffic accident description unit.

[0038] In an alternative implementation, based on the traffic accident description features of the traffic collection image stream data and the traffic accident description features corresponding to each sample traffic accident data, x traffic accident data associated with the traffic collection image stream data are determined from each sample traffic accident data, including: outputting the first x traffic accident data that best match each sample traffic accident data with the traffic collection image stream data as x traffic accident data, where x is a positive integer.

[0039] In an alternative embodiment, a first correlation of the sample traffic accident data is determined based on the traffic accident description features of a sample traffic accident data and the traffic accident description features of the traffic collected image stream data. The first correlation is used to measure the feature distance between the traffic accident description features of the sample traffic accident data and the traffic accident description features of the traffic collected image stream data.

[0040] Step S140, determining one or more traffic accident data from the x traffic accident data to be output as second-category traffic accident data corresponding to the traffic collection image stream data, where the second-category traffic accident data represents traffic accident data that does not correspond to the traffic collection image stream data.

[0041] In an alternative implementation, the second type of traffic accident data corresponding to the traffic collection image stream data represents traffic accident data that does not correspond to the traffic collection image stream data. The second type of traffic accident data can be divided into simple second type of traffic accident data and noise second type of traffic accident data. The simple second type of traffic accident data represents second type of traffic accident data that is completely irrelevant to the traffic collection image stream data, and the noise second type of traffic accident data represents second type of traffic accident data that is partially relevant to the traffic collection image stream data.

[0042] In an alternative implementation, the second-category traffic accident data determined in step S140 belongs to noise second-category traffic accident data.

[0043] Step S150, training a traffic accident recognition network based on the first type of traffic accident data and the second type of traffic accident data corresponding to the traffic collection image stream data to generate a target traffic accident recognition network, wherein the target traffic accident recognition network is used to recognize traffic accident data existing in any collected traffic collection image stream data.

[0044] In an alternative implementation, the weight information of the traffic accident recognition network includes: the weight information of the first traffic accident description unit and the weight information of the second traffic accident description unit. In an alternative implementation, training the traffic accident recognition network includes: fixing the weight information of the second traffic accident description unit and training the first traffic accident description unit.

[0045] In an alternative embodiment, the traffic accident recognition network further includes a fully connected output unit, which is used to determine the feature distance between the traffic collection image stream data and the traffic accident data based on the traffic accident description features of the traffic collection image stream data and the traffic accident description features of the traffic accident data. Thus, based on the first type of traffic accident data and the second type of traffic accident data corresponding to the traffic collection image stream data, the weight information of the first traffic accident description unit and the fully connected output unit in the traffic accident recognition network can be trained to generate a target traffic accident recognition network.

[0046] In an alternative embodiment, the knowledge learning of the traffic accident recognition network includes one or more knowledge learning stages. In each of the one or more knowledge learning stages, M traffic collection image stream data are selected, and the first traffic accident description unit in the knowledge learning stage is used to calculate the traffic accident description features corresponding to the M traffic collection image stream data respectively. For any one of the M traffic collection image stream data, a traffic accident description is performed on the traffic collection image stream data according to the first traffic accident description unit in the knowledge learning stage to generate the traffic accident description features of the traffic collection image stream data; based on the traffic accident description features of the traffic collection image stream data and the traffic accident description features corresponding to one or more sample traffic accident data respectively, x traffic accident data are selected, and one or more second-category traffic accident data corresponding to the traffic collection image stream data are determined from the x traffic accident data.

[0047] In an alternative implementation, the network learning error value of the traffic accident recognition network is calculated based on the traffic accident description features of the traffic collection image stream data and the traffic accident description features of the traffic accident data. When the network learning error value has not converged, the weight information of the traffic accident recognition network can be iteratively updated to generate a target traffic accident recognition network, and the target traffic accident recognition network is subjected to the next knowledge learning stage; when the network learning error value converges, the weight information of the traffic accident recognition network can be updated to generate a target traffic accident recognition network, and the knowledge learning of the traffic accident recognition network is terminated.

[0048] Therefore, in the process of knowledge learning of the traffic accident recognition network, one or more second-category traffic accident data corresponding to the traffic collection image stream data are determined based on the feature distance between the traffic accident description features of the traffic collection image stream data and the traffic accident description features of the sample traffic accident data. Compared with manually calibrating the second-category traffic accident data corresponding to the traffic collection image stream data, the collection speed of the second-category traffic accident data is improved, thereby contributing more second-category traffic accident data to the knowledge learning of the traffic accident recognition network, so that the traffic accident recognition network can effectively learn the difference between the traffic collection image stream data and the second-category traffic accident data, improve the observation performance of the traffic accident recognition network on the second-category traffic accident data, and then improve the traffic accident recognition accuracy of the target traffic accident recognition network.

[0049] The following is an explanation of the method for determining the second type of traffic accident data.

[0050] In an alternative implementation, one or more traffic accident data can be determined from x traffic accident data to be output as second-category traffic accident data corresponding to the traffic collection image stream data, including: from the traffic accident data other than one or more first-category traffic accident data included in the x traffic accident data, one or more traffic accident data can be determined to be output as second-category traffic accident data corresponding to the traffic collection image stream data.

[0051] Therefore, the x traffic accident data are the first x traffic accident data with the smallest feature distance to the traffic collection image stream data among one or more sample traffic accident data, and the x traffic accident data may contain the first type of traffic accident data corresponding to the traffic collection image stream data. On this basis, it is necessary to remove the first type of traffic accident data corresponding to the traffic collection image stream data from the x traffic accident data to avoid mistakenly identifying the first type of traffic accident data as the second type of traffic accident data, so that the positive sample learning data and the negative sample learning data are the same, and the interference caused by the sample learning data is introduced into the knowledge learning of the traffic accident recognition network, thereby improving the efficiency of the traffic accident recognition network in knowledge learning.

[0052] Assuming that x is equal to 5, for a certain traffic collection image stream data A, the x traffic accident data include: traffic accident data a, traffic accident data b, traffic accident data c, traffic accident data d and traffic accident data e, and traffic accident data a is the first type of traffic accident data of the traffic collection image stream data A. Then, from the traffic accident data b, traffic accident data c, traffic accident data d and traffic accident data e, determine one or more traffic accident data outputs as the second type of traffic accident data corresponding to the traffic collection image stream data A.

[0053] In an alternative implementation, when recording the first type of traffic accident data (or the traffic accident description features of the first type of traffic accident data), the image stream data ID corresponding to the first type of traffic accident data can be recorded synchronously, and the image stream data ID can point to the traffic collection image stream data. After determining the x traffic accident data corresponding to a certain traffic collection image stream data, it is judged whether the image stream data ID corresponding to the x traffic accident data is the image stream data ID of the traffic collection image stream data; if the ID tracing information corresponding to a certain traffic accident data is the same as the image stream data ID of the traffic collection image stream data among the x traffic accident data, it means that the traffic accident data is the first type of traffic accident data corresponding to the traffic collection image stream data, and the traffic accident data is removed from the x traffic accident data; if the ID tracing information corresponding to a certain traffic accident data is different from the image stream data ID of the traffic collection image stream data among the x traffic accident data, it means that the traffic accident data is not the first type of traffic accident data corresponding to the traffic collection image stream data, and the traffic accident data is retained in the x traffic accident data.

[0054] In an alternative implementation, one or more traffic accident data are randomly selected from the traffic accident data other than one or more first-category traffic accident data included in the x traffic accident data and output as the second-category traffic accident data corresponding to the traffic collection image stream data. For example, 10 traffic accident data are randomly selected from the traffic accident data other than one or more first-category traffic accident data included in the x traffic accident data and output as the second-category traffic accident data corresponding to the traffic collection image stream data.

[0055] For example, in the same knowledge learning stage, the number of second-category traffic accident data corresponding to any traffic collection image stream data is the same. r second-category traffic accident data can be determined from the x traffic accident data corresponding to "traffic collection image stream data 1", and r second-category traffic accident data can be determined from the x traffic accident data corresponding to "traffic collection image stream data 2", where r is a positive integer less than x.

[0056] In an alternative embodiment, a traffic accident description feature of the traffic collection image stream data is determined based on a first traffic accident description unit, and x traffic accident data are selected based on the traffic accident description feature of the traffic collection image stream data and the traffic accident description features of one or more sample traffic accident data included in a traffic accident database. By removing the first type of traffic accident data and pseudo second type of traffic accident data from the x traffic accident data, multiple second type of traffic accident data corresponding to the traffic collection image stream data can be obtained. Using one first type of traffic accident data and multiple second type of traffic accident data corresponding to the traffic collection image stream data, the traffic accident recognition network is updated to generate updated network weight information.

[0057] For a certain traffic collection image stream data, the characteristic distance between the traffic collection image stream data and the prior labeled traffic accident data 1 is 0.935, the characteristic distance between the traffic collection image stream data and the prior labeled traffic accident data 2 is equal to 0.935, the characteristic distance between the traffic collection image stream data and the prior labeled traffic accident data 3 is equal to 0.945, and the characteristic distance between the traffic collection image stream data and the prior labeled traffic accident data 4 is equal to 0.92; assuming that there are various restrictions in the knowledge learning of the traffic accident recognition network, each traffic collection image stream data corresponds to at most two first-class traffic accident data, and the prior labeled traffic accident data 1 and the prior labeled traffic accident data 3 can be used as the first-class traffic accident data of the traffic collection image stream data.

[0058] Therefore, although the prior labeled traffic accident data 2 and 4 have a small characteristic distance with the traffic collection image stream data, the prior labeled traffic accident data 2 and 4 are not selected as the first category traffic accident data, that is, the prior labeled traffic accident data 2 and 4 belong to the pseudo second category traffic accident data corresponding to the traffic collection image stream data.

[0059] In order to improve the reliability of the second-category traffic accident data, it is necessary to screen the traffic accident data included in the x traffic accident data except for one or more first-category traffic accident data, remove the pseudo second-category traffic accident data, and generate the second-category traffic accident data corresponding to the traffic collection image stream data.

[0060] For example, in an alternative embodiment, one or more traffic accident data are determined from the traffic accident data other than one or more first-category traffic accident data included in the x traffic accident data and output as second-category traffic accident data corresponding to the traffic collection image stream data, including: removing one or more first-category traffic accident data from the x traffic accident data to generate one or more remaining traffic accident data; based on the feature distance between the traffic collection image stream data and each of the remaining traffic accident data, determining one or more traffic accident data with the largest feature distance with the traffic collection image stream data; and outputting the one or more traffic accident data with the largest feature distance as second-category traffic accident data corresponding to the traffic collection image stream data.

[0061] In an alternative implementation, the traffic accident data with the largest feature distance represents: one or more traffic accident data with the smallest correlation with the traffic collection image stream data among the remaining traffic accident data. That is, for any second-category traffic accident data, the feature distance between the traffic accident description feature of the second-category traffic accident data and the traffic accident description feature of the traffic collection image stream data is not less than the feature distance between the traffic accident description feature of other traffic accident data and the traffic accident description feature of the traffic collection image stream data, and the other traffic accident data represents: any traffic accident data that has not been selected among the remaining traffic accident data.

[0062] In an alternative implementation, in step S130, after the x traffic accident data are determined, the x traffic accident data are sorted based on the characteristic distance between the x traffic accident data and the traffic collection image stream data to generate a traffic accident data sequence.

[0063] In another alternative embodiment, one or more traffic accident data are determined from the x traffic accident data except one or more first-category traffic accident data to be output as second-category traffic accident data corresponding to the traffic collection image stream data, including: based on the characteristic distance between the traffic collection image stream data and each of the remaining traffic accident data, determining y traffic accident data having the largest characteristic distance with the traffic collection image stream data, where y is a positive integer less than x; removing one or more first-category traffic accident data from the y traffic accident data, and outputting the remaining traffic accident data as second-category traffic accident data corresponding to the traffic collection image stream data.

[0064] The following further introduces an embodiment of knowledge learning of the traffic accident identification network.

[0065] In an alternative implementation, step S150: training a traffic accident recognition network based on the first type of traffic accident data and the second type of traffic accident data corresponding to the traffic collection image stream data to generate a target traffic accident recognition network includes the following steps.

[0066] Sub-step S152: Based on the first type of traffic accident data and the second type of traffic accident data corresponding to the traffic collection image stream data, determine the network learning error value of the traffic accident recognition network, and the network learning error value reflects the effectiveness of the traffic accident recognition network in identifying whether the traffic accident data is the first type of traffic accident data or the second type of traffic accident data.

[0067] In an alternative implementation, the network learning error value may be a cross entropy loss function value.

[0068] In an alternative embodiment, sub-step S152 includes the following steps: for the target traffic accident data, based on the feature distance between the traffic accident description features of the traffic collection image stream data and the traffic accident description features of the target traffic accident data, and the traffic accident type information of the target traffic accident data, determine the network learning error value of the traffic accident recognition network, the target traffic accident data is any first-class traffic accident data or any second-class traffic accident data corresponding to the traffic collection image stream data, and the traffic accident type information reflects that the target traffic accident data belongs to the first-class traffic accident data or the second-class traffic accident data.

[0069] In an alternative implementation, for a certain traffic collection image stream data, the traffic accident data corresponding to the traffic collection image stream data includes a first type of traffic accident data and one or more second type of traffic accident data. The target traffic accident data represents the first type of traffic accident data or any one of the one or more second type of traffic accident data.

[0070] For example, for any traffic collection image stream data in a knowledge learning stage, the traffic collection image stream data corresponds to 2 first-category traffic accident data and 58 second-category traffic accident data, that is, a total of 60 traffic accident data related to the traffic collection image stream data participate in the network learning error value determination process.

[0071] In an alternative implementation, the above-mentioned one first-category traffic accident data represents the first-category traffic accident data with the lowest characteristic distance to the traffic collection image stream data among one or more first-category traffic accident data. In other words, in a knowledge learning stage, for any traffic collection image stream data, one first-category traffic accident data is determined from one or more first-category traffic accident data corresponding to the traffic collection image stream data, and the selected first-category traffic accident data participates in the knowledge learning stage, while other first-category traffic accident data corresponding to the traffic collection image stream data do not participate in the knowledge learning stage.

[0072] In each knowledge learning stage, one or more first-class traffic accident data corresponding to the traffic collection image stream data is selected from the one or more first-class traffic accident data, and the first-class traffic accident data is used to participate in the network learning error value calculation. One or more first-class traffic accident data other than the first-class traffic accident data do not participate in the knowledge learning of this round, and the network learning error value calculation.

[0073] Sub-step S155: Based on the network learning error value, the first traffic accident description unit is trained to generate a target traffic accident recognition network, and the weight information of the second traffic accident description unit is locked.

[0074] In an alternative implementation, after training the first traffic accident description unit based on the network learning error value and generating the target traffic accident recognition network, it also includes: if the traffic accident recognition network does not meet the network convergence requirements, continuing to iteratively update the target traffic accident recognition network, and again starting from the step of acquiring the traffic collection image stream data. For the same traffic collection image stream data, the traffic accident description features of the traffic collection image stream data are different in different batches of network learning processes, and the second type of traffic accident data corresponding to the traffic collection image stream data are different in different batches of network learning processes.

[0075] The following introduces the further knowledge learning process of the traffic accident recognition network.

[0076] The knowledge learning A of a traffic accident recognition network includes the following steps.

[0077] Step D10, acquiring traffic collection image stream data and one or more first-category traffic accident data corresponding to the traffic collection image stream data, wherein the first-category traffic accident data represents traffic accident data corresponding to the traffic collection image stream data.

[0078] Step D20: for any one of the x traffic collection image stream data, extract the traffic accident description feature of the traffic collection image stream data according to the first traffic accident description unit.

[0079] Step D30, based on the traffic accident description features of the traffic collection image stream data and the traffic accident description features corresponding to each sample traffic accident data, x traffic accident data associated with the traffic collection image stream data are determined from each sample traffic accident data.

[0080] Step D40, remove one or more first-category traffic accident data from the x traffic accident data to generate one or more remaining traffic accident data; based on the characteristic distance between the traffic collection image stream data and each of the remaining traffic accident data, determine one or more traffic accident data with the largest characteristic distance from the traffic collection image stream data; output the one or more traffic accident data with the largest characteristic distance as the second-category traffic accident data corresponding to the traffic collection image stream data.

[0081] Step D50, for the target traffic accident data, based on the feature distance between the traffic accident description features of the traffic acquisition image stream data and the traffic accident description features of the target traffic accident data, and the traffic accident type information of the target traffic accident data, determine the network learning error value of the traffic accident recognition network.

[0082] Step D60: training the first traffic accident description unit based on the network learning error value to generate a target traffic accident recognition network.

[0083] Step D70, determining whether the traffic accident recognition network has converged, and when it has converged (such as when the network learning error value no longer continues to decrease), a target traffic accident recognition network is obtained.

[0084] The following describes the process of finding x traffic accident data from sample traffic accident data.

[0085] In an alternative implementation, step S130 determines x traffic accident data associated with the traffic collection image stream data from each sample traffic accident data based on the traffic accident description features of the traffic collection image stream data and the traffic accident description features corresponding to each sample traffic accident data.

[0086] In an alternative implementation, step S130 may include the following sub-steps: Sub-step S131, generates a traffic accident reference sequence based on the traffic accident description features of the sample traffic accident data in the upper layer of the traffic accident identification bitmap. The traffic accident identification bitmap is used to record the traffic accident description features corresponding to one or more sample traffic accident data based on the feature distance between multiple sample traffic accident data. The traffic accident reference sequence is used to dynamically record the identification bit information of the sample traffic accident data that is associated with the traffic collection image stream data in the traffic accident identification bitmap during the traffic accident data selection process.

[0087] In an alternative implementation, the traffic accident description feature of the sample traffic accident data in the upper layer of the traffic accident identification bitmap represents the first traffic accident description feature in the traffic accident identification bitmap that has a feature distance calculated with the traffic collection image stream data. For any traffic collection image stream data, in the process of determining x traffic accident data, it is necessary to calculate the feature distance with the sample traffic accident data in the upper layer of the traffic accident identification bitmap.

[0088] In sub-step S133, for the zth identification information included in the traffic accident reference sequence, one or more neighbor identification information associated with the zth identification information in the traffic accident identification bitmap is added to the traffic accident reference sequence, and the traffic accident description features of the sample traffic accident data corresponding to the neighbor identification information are adjacent to the traffic accident description features of the sample traffic accident data corresponding to the zth identification information, and z is a positive integer.

[0089] Sub-step S135, determines x associated identification information from the traffic accident reference sequence to obtain an adjusted traffic accident reference sequence; the feature distance between the traffic accident description features of the sample traffic accident data corresponding to the associated identification information and the traffic accident description features of the traffic collection image stream data is higher than the feature distance between the traffic accident description features of the sample traffic accident data corresponding to other identification information not selected in the traffic accident reference sequence and the traffic accident description features of the traffic collection image stream data.

[0090] Sub-step S137, again starting from the step of adding one or more neighbor identification bit information associated with the z-th identification bit information in the traffic accident identification bitmap to the traffic accident reference sequence, and executing the subsequent determination process.

[0091] Sub-step S139, if the adjusted traffic accident reference sequence is the same as the adjusted traffic accident reference sequence obtained in the subsequent determination process, the sample traffic accident data corresponding to the x identification bit information included in the adjusted traffic accident reference sequence are used as x traffic accident data.

[0092] The following describes the implementation method for establishing a traffic accident identification bitmap.

[0093] In an alternative implementation, the above method also includes: for any one sample traffic accident data among one or more sample traffic accident data, outputting the sample traffic accident data as pending traffic accident data; determining the traffic accident description features of the pending traffic accident data based on the second traffic accident description unit; determining the first identification bit information of the traffic accident description features of the pending traffic accident data in the traffic accident identification bitmap based on the traffic accident description features of the pending traffic accident data and the traffic accident description features of the traffic accident data recorded in the traffic accident identification bitmap, and the recorded traffic accident data belongs to the sample traffic accident data; determining the k recorded traffic accident data with the shortest distance to the features of the pending traffic accident data, and using the identification bit information corresponding to the k recorded traffic accident data as the neighbor identification bit information associated with the first identification bit information, where k is a positive integer.

[0094] In an alternative implementation, before the traffic accident recognition network performs knowledge learning, a traffic accident identification bitmap is established. The traffic accident description features are added to the traffic accident identification bitmap based on the feature distance between the traffic accident description features.

[0095] The following describes how to obtain the first type of traffic accident data.

[0096] Obtaining traffic collection image stream data, and one or more first-category traffic accident data corresponding to the traffic collection image stream data, includes: obtaining traffic collection image stream data, and multiple prior-annotated traffic accident data corresponding to the traffic collection image stream data; extracting one or more prior-annotated traffic accident data whose traffic accident degree is not less than a first threshold value from the multiple prior-annotated traffic accident data as the one or more first-category traffic accident data corresponding to the traffic collection image stream data.

[0097] In an alternative embodiment, one or more prior annotated traffic accident data whose degree of traffic accident is not less than a first threshold value are extracted from a plurality of prior annotated traffic accident data as one or more first-category traffic accident data corresponding to the traffic collection image stream data, including: extracting one or more prior annotated traffic accident data whose degree of traffic accident is not less than a first threshold value from a plurality of prior annotated traffic accident data; based on the accuracy information respectively corresponding to the one or more prior annotated traffic accident data, selecting one or more first-category traffic accident data corresponding to the traffic collection image stream data from the one or more prior annotated traffic accident data.

[0098] Figure 2 The traffic accident identification system 100 based on image processing which can be used to implement various embodiments described in this application is schematically shown.

[0099] For one embodiment, Figure 2A traffic accident recognition system 100 based on image processing is shown, which has one or more processors 102, a control module (chip set) 104 coupled to one or more of the (one or more) processors 102, a memory 106 coupled to the control module 104, a non-volatile memory (NVM) / storage device 108 coupled to the control module 104, one or more input / output devices 110 coupled to the control module 104, and a network interface 112 coupled to the control module 106.

[0100] The processor 102 may include one or more single-core or multi-core processors, and the processor 102 may include any combination of general-purpose processors or special-purpose processors (such as graphics processors, application processors, baseband processors, etc.). In some exemplary design ideas, the traffic accident recognition system 100 based on image processing can be used as a server device such as a gateway described in the embodiments of the present application.

[0101] In some exemplary design concepts, the image processing-based traffic accident identification system 100 may include one or more computer-readable media (e.g., memory 106 or NVM / storage device 108) having instructions 114 and one or more processors 102 integrated with the one or more computer-readable media and configured to execute instructions 114 to implement a module to perform the actions described in the present disclosure.

[0102] For one embodiment, the control module 104 may include any suitable interface controller to provide any suitable interface to one or more of the processor(s) 102 and / or any suitable device or component in communication with the control module 104 .

[0103] The control module 104 may include a memory controller module to provide an interface to the memory 106. The memory controller module may be a hardware module, a software module, and / or a firmware module.

[0104] The memory 106 may be used, for example, to load and store data and / or instructions 114 for the image processing-based traffic accident recognition system 100. For one embodiment, the memory 106 may include any suitable volatile memory, for example, a suitable DRAM. In some exemplary design ideas, the memory 106 may include a double data rate type four synchronous dynamic random access memory (DDR4 SDRAM).

[0105] For one embodiment, control module 104 may include one or more input / output controllers to provide interfaces to NVM / storage device 108 and input / output device(s) 110 .

[0106] For example, NVM / storage 108 may be used to store data and / or instructions 114. NVM / storage 108 may include any suitable non-volatile memory (e.g., flash memory) and / or may include any suitable non-volatile storage device(s) (e.g., one or more hard disk drives (HDDs), one or more compact disk (CD) drives, and / or one or more digital versatile disk (DVD) drives).

[0107] The NVM / storage device 108 may include storage resources that are physically part of the device on which the image processing-based traffic accident identification system 100 is installed, or it may be accessible to the device without being part of the device. For example, the NVM / storage device 108 may be accessed via (one or more) input / output devices 110 based on a network.

[0108] The (one or more) input / output devices 110 may provide an interface for the image processing-based traffic accident identification system 100 to communicate with any other appropriate device, and the input / output device 110 may include a communication component, a pinyin component, a sensor component, etc. The network interface 112 may provide an interface for the image processing-based traffic accident identification system 100 to communicate based on one or more networks, and the image processing-based traffic accident identification system 100 may wirelessly communicate with one or more components of a wireless network based on any of one or more wireless network standards and / or protocols, for example, access a wireless network based on a communication standard, such as WwFw, 2G, 3G, 4G, 5G, etc., or a combination thereof for wireless communication.

[0109] For one embodiment, one or more of the processor(s) 102 may be loaded together with the logic of one or more controllers (e.g., a memory controller module) of the control module 104. For one embodiment, one or more of the processor(s) 102 may be loaded together with the logic of one or more controllers of the control module 104 to form a system level load. For one embodiment, one or more of the processor(s) 102 may be integrated on the same die with the logic of one or more controllers of the control module 104. For one embodiment, one or more of the processor(s) 102 may be integrated on the same die with the logic of one or more controllers of the control module 104 to form a system on chip (SoC).

[0110] In various embodiments, the image processing-based traffic accident identification system 100 may be, but is not limited to: an image processing-based traffic accident identification system, a desktop computing device or a mobile computing device (e.g., a laptop computing device, a handheld computing device, a tablet computer, a netbook, etc.) and other terminal devices. In various embodiments, the image processing-based traffic accident identification system 100 may have more or fewer components and / or different architectures. For example, in some exemplary design ideas, the image processing-based traffic accident identification system 100 includes one or more cameras, a keyboard, a liquid crystal display (LCD) screen (including a touch screen display), a non-volatile memory port, multiple antennas, a graphics chip, an application-specific integrated circuit (ASIC), and a speaker.

[0111] The embodiments of the present application are introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core idea of ​​the present application. At the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A traffic accident recognition method based on image processing, characterized in that: The method is applied to a traffic accident recognition system based on image processing and is executed by a traffic accident recognition network, wherein the traffic accident recognition network includes a first traffic accident description unit and a second traffic accident description unit. The method includes: Acquire traffic collection image stream data, and one or more first-category traffic accident data corresponding to the traffic collection image stream data, wherein the first-category traffic accident data represents traffic accident data corresponding to the traffic collection image stream data; Extracting traffic accident description features of the traffic collection image stream data according to the first traffic accident description unit, the traffic accident description features including trajectory feature vectors of dynamic moving elements and trajectory feature vectors of accident-related elements in the traffic collection image stream data, the trajectory feature vectors of the dynamic moving elements and the trajectory feature vectors of the accident-related elements at least including position information of the dynamic moving elements and the accident-related elements; Based on the traffic accident description features of the traffic collection image stream data and the traffic accident description features corresponding to each sample traffic accident data, x traffic accident data associated with the traffic collection image stream data are determined from each sample traffic accident data, the traffic accident description features corresponding to the sample traffic accident data are obtained based on the second traffic accident description unit, and x is a positive integer; From the x traffic accident data, determine one or more traffic accident data to be output as second-category traffic accident data corresponding to the traffic collection image stream data, wherein the second-category traffic accident data represents traffic accident data that does not correspond to the traffic collection image stream data; The traffic accident recognition network is trained based on the first type of traffic accident data and the second type of traffic accident data corresponding to the traffic collection image stream data to generate a target traffic accident recognition network, wherein the target traffic accident recognition network is used to recognize traffic accident data existing in any collected traffic collection image stream data.

2. The traffic accident identification method based on image processing according to claim 1, characterized in that: The step of determining, from the x traffic accident data, one or more traffic accident data to be output as the second type of traffic accident data corresponding to the traffic collection image stream data comprises: From the traffic accident data included in the x traffic accident data except the one or more first-category traffic accident data, one or more traffic accident data are determined to be output as second-category traffic accident data corresponding to the traffic collection image stream data.

3. The traffic accident identification method based on image processing according to claim 2 is characterized in that: The step of determining, from the traffic accident data other than the one or more first-category traffic accident data included in the x traffic accident data, one or more traffic accident data to be output as the second-category traffic accident data corresponding to the traffic collection image stream data comprises: Removing the one or more first-category traffic accident data from the x traffic accident data to generate one or more remaining traffic accident data; Based on the characteristic distances between the traffic collection image stream data and each of the remaining traffic accident data, determining one or more traffic accident data having the largest characteristic distances to the traffic collection image stream data; Outputting the one or more traffic accident data with the largest characteristic distance as the second type of traffic accident data corresponding to the traffic collection image stream data; Alternatively, based on the characteristic distances between the traffic collection image stream data and each of the remaining traffic accident data, y traffic accident data having the largest characteristic distances with the traffic collection image stream data are determined, where y is a positive integer less than x; The one or more first-category traffic accident data are removed from the y traffic accident data, and the remaining traffic accident data are output as second-category traffic accident data corresponding to the traffic collection image stream data.

4. The method for traffic accident identification based on image processing according to claim 1, characterized in that: The training of the traffic accident recognition network based on the first type of traffic accident data and the second type of traffic accident data corresponding to the traffic collection image stream data to generate a target traffic accident recognition network includes: Based on the first type of traffic accident data and the second type of traffic accident data corresponding to the traffic collection image stream data, determining a network learning error value of the traffic accident recognition network, wherein the network learning error value reflects the effectiveness of the traffic accident recognition network in identifying whether the traffic accident data is the first type of traffic accident data or the second type of traffic accident data; Based on the network learning error value, the first traffic accident description unit is trained to generate the target traffic accident recognition network, and the weight information of the second traffic accident description unit is locked.

5. The method for traffic accident identification based on image processing according to claim 4, characterized in that: The determining of the network learning error value of the traffic accident recognition network based on the first type of traffic accident data and the second type of traffic accident data corresponding to the traffic collection image stream data comprises: For target traffic accident data, determining a network learning error value of the traffic accident recognition network based on a feature distance between traffic accident description features of the traffic collection image stream data and traffic accident description features of the target traffic accident data, and traffic accident type information of the target traffic accident data; Among them, the target traffic accident data is any one of the first type of traffic accident data or any one of the second type of traffic accident data corresponding to the traffic collection image stream data, and the traffic accident type information reflects that the target traffic accident data belongs to the first type of traffic accident data or the second type of traffic accident data.

6. The method for traffic accident identification based on image processing according to claim 4, characterized in that: After the first traffic accident description unit is trained based on the network learning error value to generate the target traffic accident identification network, the method further includes: If the traffic accident recognition network does not meet the network convergence requirement, the target traffic accident recognition network is continuously iterated and updated, and the process returns to the step of obtaining the traffic collection image stream data and continues the iteration; For the same traffic collection image stream data, the traffic accident description features of the traffic collection image stream data are different in different batches of network learning processes, and the second type of traffic accident data corresponding to the traffic collection image stream data are different in different batches of network learning processes.

7. The method for traffic accident identification based on image processing according to claim 1, characterized in that: The traffic accident description features based on the traffic collection image stream data and the traffic accident description features corresponding to each sample traffic accident data, and determining x traffic accident data associated with the traffic collection image stream data from each sample traffic accident data, include: Generate a traffic accident reference sequence based on the traffic accident description features of the sample traffic accident data in the upper layer of the traffic accident identification bitmap, wherein the traffic accident identification bitmap is used to record the traffic accident description features corresponding to each sample traffic accident data based on the feature distance between the multiple sample traffic accident data, and the traffic accident reference sequence is used to dynamically record the identification bit information of the sample traffic accident data associated with the traffic collection image stream data in the traffic accident identification bitmap during the determination process of the traffic accident data; For the zth identification bit information included in the traffic accident reference sequence, one or more neighbor identification bit information associated with the zth identification bit information in the traffic accident identification bitmap is added to the traffic accident reference sequence, the traffic accident description feature of the sample traffic accident data corresponding to the neighbor identification bit information is adjacent to the traffic accident description feature of the sample traffic accident data corresponding to the zth identification bit information, and z is a positive integer; Determine x pieces of associated identification information from the traffic accident reference sequence to generate an adjusted traffic accident reference sequence; The feature distance between the traffic accident description feature of the sample traffic accident data corresponding to the associated identification information and the traffic accident description feature of the traffic collected image stream data is smaller than the feature distance between the traffic accident description feature of the sample traffic accident data corresponding to other identification information not selected in the traffic accident reference sequence and the traffic accident description feature of the traffic collected image stream data; Starting again from the step of adding one or more neighbor identification bit information associated with the z-th identification bit information in the traffic accident identification bitmap into the traffic accident reference sequence, executing a subsequent determination process; If the adjusted traffic accident reference sequence is the same as the adjusted traffic accident reference sequence obtained by the subsequent determination process, the sample traffic accident data corresponding to the x identification bit information included in the adjusted traffic accident reference sequence are used as the x traffic accident data.

8. The method for traffic accident identification based on image processing according to claim 7, characterized in that: The method further comprises: For any one of the sample traffic accident data, output the sample traffic accident data as pending traffic accident data; determining, according to the second traffic accident description unit, a traffic accident description feature of the pending traffic accident data; Based on the traffic accident description features of the pending traffic accident data and the traffic accident description features of the traffic accident data recorded in the traffic accident identification bitmap, determining the first identification bit information of the traffic accident description features of the pending traffic accident data in the traffic accident identification bitmap, and the recorded traffic accident data belongs to the sample traffic accident data; Determine k recorded traffic accident data with the smallest characteristic distance from the pending traffic accident data, and use the identification information corresponding to the k recorded traffic accident data as neighbor identification information associated with the first identification information, where k is a positive integer.

9. The method for traffic accident identification based on image processing according to claim 1, characterized in that: The acquiring of the traffic collection image stream data and one or more first-category traffic accident data corresponding to the traffic collection image stream data includes: Acquire traffic collection image stream data and a plurality of prior-labeled traffic accident data corresponding to the traffic collection image stream data; One or more a priori labeled traffic accident data with a traffic accident degree not less than a first threshold value are extracted from the multiple a priori labeled traffic accident data, and determined as one or more first-category traffic accident data corresponding to the traffic collection image stream data.

10. A traffic accident recognition system based on image processing, characterized in that: It includes a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by the processor, the traffic accident recognition method based on image processing according to any one of claims 1 to 9 is implemented.

Images