A confidence-based method for identifying non-motorized vehicles and pedestrians

Abstract

This application discloses a confidence-based method for identifying non-motorized vehicle and pedestrian individuals. It utilizes an AI-powered intelligent image review server to dynamically screen illegal images, employing a deep neural network to deconstruct the essential differences between image quality defects and missing target features, accurately filtering invalid data. Subsequently, it initiates multi-source data fusion and comparison for approved images, innovatively integrating multi-source data into an identity DNA confidence model. This model dynamically adjusts the contribution of each data source through an adaptive weight allocation mechanism, ultimately outputting a continuously quantified identity confidence score. In this way, by deeply coupling the original image quality assessment with the dynamic fusion of multi-source data, a complete confidence chain from data cleaning to hierarchical judgment is constructed, eliminating the consumption of computational resources by invalid data and achieving the ability to verify the identity confidence score of violators based on a pedestrian and non-motorized vehicle confidence model.

Description

Technical Field

[0001] This application relates to the field of intelligent recognition, and more specifically, to a confidence-based method for identifying non-motorized vehicle pedestrians. Background Technology

[0002] In the field of intelligent traffic enforcement, the identification of non-motorized vehicles and pedestrians violating traffic rules is facing increasingly severe technical challenges. With the surge in urban slow-moving traffic, the amount of violation image data captured daily by electronic eyes is growing exponentially, of which up to 42% is invalid or inefficient data—including technically unusable images due to motion blur and rain / fog interference, images lacking scene-specific features due to helmet obstruction and side-face avoidance, and redundant data captured repeatedly by multiple cameras. Directly feeding this raw data into the identification system would not only result in a serious waste of computing resources but also trigger a chain reaction of delays in subsequent comparison processes.

[0003] Traditional solutions typically extract and match features directly from raw images of violations. However, these methods have significant limitations: existing solutions often neglect quality screening and content verification of raw image data, and lack a systematic filtering mechanism for issues such as image clarity, duplicate submissions, and target misidentification. This results in low-quality or invalid data directly entering the recognition process, increasing subsequent computational burden and increasing the likelihood of feature extraction bias, thus reducing overall recognition reliability. Furthermore, the feature maps of violations generated by conventional visual feature extraction methods (such as dense convolution) contain a large amount of non-critical information and lack dynamic evaluation of channel-level saliency. Feature redundancy not only reduces model efficiency but also introduces noise interference, especially in complex backgrounds or target occlusion scenarios, easily leading to misjudgments of violations and non-motorized vehicles. More importantly, existing identity recognition methods largely rely on single-person-certificate image comparisons or basic database queries, failing to integrate multi-dimensional clues such as the spatiotemporal correlation of violation locations, continuity of historical behavior, and consistency of registration information. The one-sidedness of a single data source makes it difficult to address the challenge of ambiguous identities of violations in real-world scenarios (such as occlusion or cross-regional behavior), leading to questionable confidence in the recognition results. These limitations lead to systemic defects in the identity recognition process, such as low efficiency, poor fault tolerance, waste of resources, and insufficient credibility of results, which seriously restrict the level of automation and the improvement of law enforcement efficiency in the governance of non-motorized vehicle violations.

[0004] Therefore, an optimized confidence-based non-motorized vehicle and pedestrian identification scheme is needed. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a confidence-based method for identifying non-motorized vehicle and pedestrian identities. It utilizes an AI-powered intelligent image review server to dynamically screen illegal images, employing a deep neural network to deconstruct the essential differences between image quality defects and missing target features, accurately filtering invalid data. Subsequently, multi-source data fusion and comparison are initiated for approved images, innovatively integrating multi-source data into an identity DNA confidence model. This model dynamically adjusts the contribution of each data source through an adaptive weight allocation mechanism, ultimately outputting a continuously quantified identity confidence score. In this way, by deeply coupling the original image quality assessment with the dynamic fusion of multi-source data, a complete confidence chain from data cleaning to hierarchical judgment is constructed, eliminating the consumption of computing resources by invalid data. This enables the verification of illegal personnel identity confidence scores based on a pedestrian and non-motorized vehicle confidence model, effectively reducing the workload of police officers involved in the verification process, increasing the rate of public education and punishment, and promoting the reform of the duty mechanism.

[0006] According to one aspect of this application, a confidence-based method for identifying non-motorized pedestrians is provided, comprising:

[0007] Obtain image data of non-motorized vehicle violations collected by cameras;

[0008] The non-motorized vehicle violation image data is input into the AI ​​intelligent image review server to obtain the image review results;

[0009] In response to the review result being approved, multi-source data fusion and comparison are performed based on the non-motorized vehicle violation image data to obtain multi-source basic comparison result data;

[0010] The multi-source basic comparison results are input into the multi-source data fusion identity DNA confidence model to obtain personnel identity confidence score data.

[0011] Based on the aforementioned personnel identity confidence scores, a differentiated processing strategy is determined.

[0012] Compared with existing technologies, this application provides a confidence-based method for identifying non-motorized vehicle and pedestrian identities. It utilizes an AI-powered intelligent image review server to dynamically screen illegal images, employing deep neural networks to deconstruct the essential differences between image quality defects and missing target features, accurately filtering invalid data. Subsequently, it initiates multi-source data fusion and comparison for approved images, innovatively integrating multi-source data into an identity DNA confidence model. This model dynamically adjusts the contribution of each data source through an adaptive weight allocation mechanism, ultimately outputting a continuously quantified identity confidence score. By deeply coupling the original image quality assessment with the dynamic fusion of multi-source data, a complete confidence chain from data cleaning to hierarchical judgment is constructed, eliminating the consumption of computing resources by invalid data. This enables the verification of illegal personnel's identity confidence scores based on a pedestrian and non-motorized vehicle confidence model, effectively reducing the workload of police officers involved in the verification process, increasing the rate of public education and punishment, and promoting the reform of the duty mechanism. Attached Figure Description

[0013] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0014] Figure 1 This is a flowchart of a confidence-based non-motorized vehicle pedestrian identification method according to an embodiment of this application;

[0015] Figure 2 This is a schematic diagram of the data flow of the confidence-based non-motorized vehicle and pedestrian identification method according to an embodiment of this application;

[0016] Figure 3 This is a flowchart of sub-step S2 of the confidence-based non-motorized vehicle pedestrian identification method according to an embodiment of this application;

[0017] Figure 4 This is a schematic diagram illustrating the calculation of points for non-motorized vehicle violation candidates according to an embodiment of this application. Detailed Implementation

[0018] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0019] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0020] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0021] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0022] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0023] The technical solution of this application proposes a method for identifying non-motorized pedestrians based on confidence level. Figure 1 This is a flowchart of a confidence-based non-motorized vehicle and pedestrian identification method according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in the confidence-based non-motorized vehicle and pedestrian identification method according to an embodiment of this application. Figure 1 and Figure 2 As shown, the confidence-based non-motorized vehicle and pedestrian identification method according to an embodiment of this application includes the following steps: S1, acquiring non-motorized vehicle violation image data collected by a camera; S2, inputting the non-motorized vehicle violation image data into an AI intelligent image review server to obtain an image review result; S3, in response to the image review result being approved, performing multi-source data fusion comparison based on the non-motorized vehicle violation image data to obtain multi-source basic comparison result data; S4, inputting the multi-source basic comparison result data into a multi-source data fusion identity DNA confidence model to obtain personnel identity confidence score data; S5, determining a differentiated processing strategy based on the personnel identity confidence score data.

[0024] Specifically, S1 involves acquiring non-motorized vehicle violation image data collected by cameras. In a specific example of this application, a dual-track image data acquisition strategy is first adopted: on the one hand, dedicated non-motorized vehicle violation capture equipment is deployed to accurately capture images of violations such as running red lights and driving against traffic; on the other hand, the existing security monitoring resources widely distributed in urban intersections and road sections by relevant management systems are fully utilized; then, intelligent algorithms are used to perform real-time analysis or back-end secondary processing on these monitoring video streams to identify and extract key frame images containing non-motorized vehicle violations. This multi-source, intelligent image acquisition method lays a solid data foundation for subsequent AI review and filtering and confidence model analysis based on multi-source data, and is the primary step in solving core problems such as large workload and low accuracy in the review process.

[0025] Specifically, in step S2, the non-motorized vehicle violation image data is input into an AI intelligent image review server to obtain the review results. That is, the AI ​​intelligent image review server automatically and intelligently cleans, filters, and classifies the massive amounts of raw violation image data collected from the front end, using technical means to pre-emptively remove invalid information, significantly improving the efficiency and accuracy of subsequent identity recognition and processing.

[0026] It is worth mentioning that the AI ​​intelligent image review server is a dedicated computing system that integrates computer vision (especially object detection, behavior recognition, and face quality assessment) and rule engine. It is responsible for performing specific image analysis, filtering, and classification tasks.

[0027] During this process, the AI-powered intelligent image review server receives image data from two sources: first, images of violations directly captured by dedicated non-motorized vehicle capture equipment at the front end; and second, images captured by existing security monitoring points. All images are then fed into the non-motorized vehicle control platform and uniformly input into the AI-powered image review server. The AI-powered intelligent image review server uses a deployed deep learning model and pre-set filtering logic to perform secondary analysis and filtering of the violation images. Combined with platform logic filtering, it effectively filters out violation identification errors (errors in identifying illegal behavior or illegal object), poor image quality (blurry images from front-end capture data, data without faces), and duplicate data (multiple pushes of the same type of violation data for the same object at the same time). Based on the analysis results, the server automatically categorizes and labels the images and outputs them to three different libraries: valid and usable "approved" images enter library A; images with questionable quality or identification enter library B for subsequent reference or model training; and invalid, discarded images (such as severely blurry, images without faces, duplicates, or obvious identification errors) are directly discarded and placed in library C. The database is divided into three categories: Database A (approved database) contains high-quality, valid illegal data that proceeds to the next stage of the process; Database B (suspected database) contains boundary cases or training materials; and Database C (discarded database) contains clearly invalid data that is discarded. This classification forms the basis for differentiated processing and optimized resource allocation.

[0028] In a specific example of this application, such as Figure 3 As shown, S2 includes: S21, the AI ​​intelligent image review server performs illegal behavior identification and illegal object identification on the non-motorized vehicle violation image data to determine whether there is an illegal identification error; S22, the AI ​​intelligent image review server performs image quality assessment on the non-motorized vehicle violation image data to determine whether the image quality of the non-motorized vehicle violation image data meets preset requirements; S23, the AI ​​intelligent image review server determines whether the non-motorized vehicle violation image data is duplicate data.

[0029] Specifically, in step S21, the AI ​​intelligent image review server performs illegal behavior identification and illegal object identification on the non-motorized vehicle violation image data to determine whether there are any violations identification errors. It should be understood that in actual scenarios of non-motorized vehicle violation identification, the original captured images often lead to factual misjudgments due to environmental interference or the complexity of the target behavior. For example, normal yielding behavior may be mistakenly labeled as illegal lane changing, pedestrians may be incorrectly identified as non-motorized vehicle subjects, or the characteristics of motorized vehicles and electric bicycles may be confused. If such fundamental errors are not pre-screened during the image review stage, subsequent identity recognition processes will be based on erroneous premises, resulting in wasted system resources and deviations in legal fact determination. Therefore, a pre-emptive violation fact verification mechanism needs to be established. In the technical solution of this application, by performing illegal behavior identification and illegal object identification on the non-motorized vehicle violation image data, the authenticity of the entity and behavior of the violation event can be ensured. Through the collaborative operation of two-level identification, the system can filter out two types of key erroneous data: one is "false violations" caused by behavioral misjudgment (such as temporary lane crossing being mistakenly judged as going against traffic), and the other is "subject errors" caused by object confusion (such as pedestrians pushing bicycles being mistakenly identified as cyclists). This prevents invalid data from flowing into subsequent processes at the source, while providing behaviorally calibrated target entity features for multi-source comparison.

[0030] Specifically, firstly, visual feature extraction of the non-motorized vehicle violation image data is performed based on convolutional coding to obtain a visual feature map of the violation object. It should be understood that the original non-motorized vehicle violation image data is often affected by complex environmental interference (such as changes in lighting, target occlusion, and background clutter), making it difficult to stably represent the essential characteristics of the violation object using pixel-level data directly. Traditional manual rule-based or shallow feature extraction methods are easily affected by noise and cannot adaptively capture the key component structures (such as wheel shape and handlebar outline) and behavioral characteristics (such as riding posture) of non-motorized vehicles. Therefore, in the technical solution of this application, visual feature extraction of the violation object is performed based on convolutional coding on the non-motorized vehicle violation image data. Through the hierarchical perception mechanism of convolutional neural networks, the essential visual pattern of the violation object is deconstructed from the original violation image to obtain a visual feature map of the violation object.

[0031] In this process, through multi-layer convolutional kernel scanning, local features (such as edges and textures) are gradually aggregated into global structural information (such as the overall outline of a non-motorized vehicle). Specifically, shallow convolutional layers extract basic geometric features (edges and corners) and identify local components of the non-motorized vehicle (such as tire arcs and handlebar straight lines). Deep convolution expands the receptive field, combines local components into an overall structure (such as the spatial relationship between "frame-seat-pedal"), and associates it with behavioral semantics (such as "front wheel crossing the stop line"). Finally, a three-dimensional tensor (channel × height × width) is output, where each channel corresponds to a visual mode (such as "helmet feature channel" and "license plate area channel"), and the spatial dimension preserves the target position information.

[0032] Next, it should be understood that in non-motorized vehicle violation recognition scenarios, the original violation images often contain complex background interference (such as obscured traffic flow, billboards), variable target object postures (such as sideways riding, helmet obstruction), and environmental noise (such as rain and fog blurring, low light at night). Although the visual feature map of the violation object extracted by conventional convolution contains target information, a large number of channels carry redundant background features or low signal-to-noise ratio responses. If such dense features are directly used for violation object judgment, the model is easily affected by irrelevant visual patterns, especially when the boundary between the violation object and the non-motorized vehicle is blurred (such as electric scooters and bicycles having similar shapes), which may lead to misjudgment due to noise accumulation. Therefore, in the technical solution of this application, the visual feature map of the violation object is subjected to visual feature sparsification to obtain a sparse coding map of the visual features of the violation object. In specific implementation, firstly, the visual feature map of the violation object is decomposed along the channel dimension, and the two-dimensional spatial features of each channel are flattened into independent vectors to form a sequence distribution of the visual feature vectors of the violation object in the channel unit; then, the long-range dependency relationship between the channel vectors is analyzed through a self-attention mechanism. For example, by capturing the correlation between the "non-motorized vehicle tire texture" and "cyclist leg posture" channels and suppressing the weights of irrelevant channels such as "roadside vegetation," a sequence distribution of the visual feature encoding vectors of the illegal object in the context of the channel unit is obtained. Furthermore, addressing the problem that traditional Shannon entropy fails during local mutations, this application first constructs an uncertain configuration space to quantify the steady-state shift during feature mutations (such as the sudden appearance of an umbrella reducing the salience of the "facial features" channel). Then, channel weights are corrected through group orbital integrals to ensure that key channels ("non-motorized vehicle outline") remain highly activated under interference. Subsequently, a binary mask is generated based on the corrected channel salience: high-salience channels (>threshold) retain the original features, while low-salience channels are set to zero. In this way, visual feature sparsity, as a pre-optimization module for illegal object recognition, solves the problems of feature interference and noise propagation in complex scenarios through entropy-driven multi-steady-state channel selection.

[0033] Specifically, firstly, it should be understood that the visual feature map of the illegal object extracted by the convolutional neural network is essentially a three-dimensional tensor (height × width × number of channels), where each channel carries a specific type of visual response pattern (such as edge, texture, and color distribution). Traditional methods, when directly manipulating the three-dimensional tensor for feature fusion, struggle to explicitly model the semantic relationships between channels. Therefore, in the technical solution of this application, the visual feature map of the illegal object is modularized along the channel dimension to obtain a sequence distribution of the visual feature vectors of the illegal object in each channel unit. This breaks the coupling between spatial location and channel response in traditional feature maps, enabling the model to independently analyze the complete spatial semantics of each channel (e.g., the distribution pattern of the "non-motorized vehicle tire" channel in the entire image). A single channel vector fully carries the global distribution of a specific visual pattern (e.g., the overall intensity of the "raincoat reflection" channel), allowing subsequent modules to directly evaluate the overall value of the pattern and avoid misjudgments caused by local interference (e.g., reflection only appears in the corners of the image).

[0034] In a specific example of this application, the visual feature map of the illegal object is modularized into feature units using the following formula:

[0035]

[0036] in, It is the visual feature map of the illegal object. It is the set of real numbers. , These are the height and width of the visual feature map of the illegal object, respectively. It is the number of channels in the visual feature map of the illegal object. It is the sequence distribution of the visual feature vectors of the illegal objects in the channel unit. These are the 1st, 2nd, and Cth visual feature vectors of the illegal object in the channel unit, respectively, in the sequence distribution of the visual feature vectors of the illegal object in the channel unit.

[0037] Next, it should be understood that although the original feature maps extracted by convolutional neural networks contain local visual elements of the target object, their channel-dimensional features are usually fragmented. Traditional methods cannot distinguish the semantic relationships between these channels, leading to chaotic feature representations for subsequent identity recognition. Therefore, in the technical solution of this application, the sequence distribution of the visual feature vectors of the illegal object in the channel unit is input into a channel feature association encoder based on the Transformer architecture to obtain the sequence distribution of the visual feature encoding vectors of the illegal object in the context of the channel unit. Through the self-attention mechanism of the Transformer, the model can dynamically capture long-distance contextual relationships across channels. In this way, the limitations of local perception of traditional convolutional features can be overcome, significantly improving the anti-interference ability of feature representation and the aggregation efficiency of identity discrimination information, providing a data foundation for subsequent accurate feature screening based on channel saliency.

[0038] In a specific example of this application, the sequence distribution of the visual feature vectors of the illegal object in the channel unit is encoded using a Transformer-based inter-channel feature correlation encoding method, as follows:

[0039]

[0040] in, Indicates context-dependent encoding. It is the sequence distribution of the visual feature encoding vectors of the illegal object associated with the channel unit context. These are the 1st, 2nd, and Cth channel unit context-associated illegal object visual feature encoding vectors in the sequence distribution of the channel unit context-associated illegal object visual feature encoding vectors.

[0041] Therefore, it should be understood that although the preceding steps have established global associations between channels through Transformer, the encoded vector sequence still contains a large amount of high-dimensional information fused across channels. Without a quantitative assessment of the importance of each channel, subsequent feature filtering will struggle to accurately retain core identity clues in complex scenarios. Therefore, in the technical solution of this application, the initial channel saliency of the context-related visual feature encoding vectors of the illegal object in the sequence distribution of the channel unit is calculated. That is, each channel unit is given an objective importance scalar assessment through a data-driven approach, establishing a quantitative benchmark for subsequent intelligent filtering of feature channels.

[0042] In practice, the system utilizes a pre-trained lightweight evaluation sub-network (such as a fully connected layer or a 1D convolutional kernel) to perform nonlinear mapping and scalar compression on the context-associated visual feature encoding vector of the illegal object for each channel unit in the sequence. The initial channel saliency score is trained end-to-end, implicitly learning the complex correlation patterns between channel information and the overall identity recognition task, making the score itself a quantified projection of the channel's contribution to the global feature representation system. This step significantly improves the early identification capability of invalid feature channels, reduces the impact of noise interference on subsequent identity assessment, and ensures the effective transmission of high-value identity features in complex law enforcement environments such as occlusion and ambiguity.

[0043] In a specific example of this application, the initial channel saliency of the context-associated illegal object visual feature encoding vector of each channel unit is calculated using the following formula:

[0044]

[0045] in, express function, , and For learnable equilibrium parameters, for Shannon entropy, for Shannon entropy and for Shannon entropy, yes The initial channel saliency of the corresponding channel unit context-associated illegal object visual feature encoding vector.

[0046] Furthermore, it should be understood that although the sequence distribution of the visual feature encoding vector of the illegal object in the context of the channel unit contains global correlation information, its Shannon entropy uncertainty measure may fall into an unsteady imbalance due to local feature mutations (such as texture cusps at the edge of the occluded area or outliers caused by rain and fog interference). Traditional balance parameter control is difficult to adapt to such complex nonlinear perturbations, resulting in a deviation in the ability of the initial channel saliency assessment to discriminate key identity features, thereby affecting the accuracy of subsequent feature selection. Therefore, in a preferred example of this application, the initial channel saliency is corrected by weighted group orbit integral correction to obtain the channel saliency. That is, by constructing a multi-stable uncertainty topological space model, the problem of assessment instability caused by mutual interference between feature channels is solved.

[0047] In practice, the Shannon entropy corresponding to the initial channel saliency is first mapped to a geometric landscape in the configuration space. Three matrix models are established to represent the baseline, interference, and differential states of the channels, and their minimum eigenvalues ​​and geometric rank norms are calculated. Then, the group orbital integral method is used to weight and integrate these matrix models, achieving a dynamic balance of multi-dimensional uncertainty distribution. In this process, the group orbital integral dynamically allocates weights within a multi-steady-state manifold space, transforming discrete channel feature perturbations into geometric morphology adjustments on a continuous manifold at the computational level. This correction mechanism based on nonlinear manifold optimization ensures that the final channel saliency retains key identity-discriminating features (such as the eye contour at the edge of the occluded area) while effectively suppressing abnormal fluctuations caused by environmental interference or local feature breaks. This results in a stable feature importance map with robustness on the edge computing device, laying the foundation for subsequent high-precision visual feature sparsity and ultimately improving the robust decision-making capability of the identity recognition system in complex law enforcement environments.

[0048] In this specific example, when using vectors respectively , and When calculating channel saliency using Shannon entropy, considering the potential local cusp abrupt changes in the sequence distribution of the channel unit context-associated illegal object visual feature encoding vector obtained after context-associated encoding via Transformer, relative to the initial sequence distribution of the channel unit illegal object visual feature vector, the uncertainty measure of Shannon entropy may also fall into a non-multistable state. This is only addressed by balancing parameters. , and It may be difficult to compensate.

[0049] Based on this, and considering the Shannon entropy's measure of uncertainty, we first introduce an uncertainty configuration space matrix to express the uncertainty distribution landscape within the steady-state configuration space:

[0050]

[0051] Furthermore, considering that multiple stable states should correspond to local minima under uncertain distribution landscapes, the matrix is ​​determined. , and The smallest eigenvalue , and However, for local cusp mutations, a low-rank representation of the geometry is required, i.e. , , Thus, the group action equivalence of multiple stable states is achieved based on the weighted group orbit integral, i.e.:

[0052]

[0053] Therefore, by nonlinear group interaction of the uncertain landscape, the multi-stable intrinsic property representation of the Shannon entropy based on uncertainty metric is performed, that is, by defining the multi-stability equilibrium position correction through the geometry of the configuration space, thereby improving the multi-equilibrium stability of the channel saliency.

[0054] Then, it should be understood that although the channel saliency map after multi-steady-state correction has accurately quantified the importance of each feature channel, the original channel feature vector sequence still contains a large amount of redundant information that interferes with identity recognition. Therefore, in the technical solution of this application, based on the channel saliency of the illegal object visual feature encoding vector of each channel unit context-associated with the channel saliency, the sequence distribution of the illegal object visual feature vector of the channel unit is sparsified to obtain the sparse sequence distribution of the illegal object visual feature vector of the channel unit.

[0055] In practice, the system iterates through the sequence distribution of visual feature vectors of illegal objects in each channel unit, setting a threshold gating strategy based on the corresponding channel saliency score. This process is algorithmically represented by a weighted mask operation. By constructing a non-linear mapping function with the channel saliency score, it achieves soft filtering and importance reweighting of feature channels. This data structure, focusing on core discriminative features, significantly improves the anti-interference capability of subsequent illegal object identification, ensuring the identity comparison module maintains high robustness even in law enforcement scenarios with missing features, ultimately supporting an improvement in the decision-making efficiency of the entire confidence assessment pipeline.

[0056] In a specific example of this application, the sequence distribution of the visual feature vector of the illegal object in the channel unit is sparsified using the following formula:

[0057]

[0058] in, Represents the sparsification function, It is a preset threshold.

[0059] Subsequently, it should be understood that although the sparse sequence generated in the preceding steps has completed the intelligent filtering of feature channels, its data structure has been transformed from a multi-dimensional feature map into a flattened sequence form. This sequence format cannot be directly connected to subsequent convolutional neural network modules. If the sequence data is directly input into the downstream model, it will lead to the loss of positional semantic information, severely restricting the model's ability to resolve cross-regional occluded targets. Therefore, in the technical solution of this application, the sparse sequence distribution of the visual feature vector of the illegal object in the channel unit is reshaped to obtain the sparse coding map of the visual features of the illegal object.

[0060] In practice, the system uses a preset spatial dimension mapping function to reverse-locate the feature vector of each channel unit in the sequence according to its row and column coordinates in the original feature map, while retaining the activation states marked by channel saliency screening (preservation or suppression). This compressed feature representation that preserves the key spatial topology allows the subsequent illegal object identification module to directly and efficiently process the feature map using standard convolutional kernels, providing structurally complete and information-density-optimized visual feature input for subsequent multi-source confidence assessment.

[0061] Finally, based on the sparse encoded map of the visual features of the violation object, it is determined whether the violation object is a non-motorized vehicle to obtain the violation object identification result. It should be understood that although the feature map after channel-level feature sparsification has focused on key visual elements, its information still lacks a semantic dimension for judging the type of violation. In a specific example of this application, the sparse encoded map of the visual features of the violation object is passed through a classifier-based violation object recognizer to obtain a classification result, which indicates whether the violation object is a non-motorized vehicle. The classifier-based violation object recognizer establishes a mapping relationship between sparse feature patterns and violation object categories through end-to-end learning. This process utilizes the high signal-to-noise ratio advantage after sparse feature compression to effectively identify the essential kinematic features of non-motorized vehicles under occlusion interference, eliminating type misjudgments caused by the visibility of local limbs. This violation subject identification based on deep semantic parsing establishes an accurate target object definition foundation for subsequent identity comparison, avoiding incorrect association of pedestrian trajectories with the non-motorized vehicle registration database, thereby improving the accuracy of type determination in complex scenarios for the entire law enforcement system and ensuring the reliability of the input source for identity confidence assessment.

[0062] Specifically, in step S22, the AI ​​intelligent image review server performs an image quality assessment on the non-motorized vehicle violation image data to determine whether the image quality of the non-motorized vehicle violation image data meets preset requirements. It should be understood that actual law enforcement images often suffer from motion blur, rain and fog interference leading to the loss of key target features, or helmet obstruction causing the loss of identifiable facial areas. For example, images of violating cyclists captured in the rain may have light spot noise due to raindrop scattering, or the helmet shadow may completely cover the facial features. If such low-quality images flow into the subsequent recognition process, it will cause feature extraction distortion and misidentification. Therefore, to construct a dynamic quality criterion system and filter invalid data from both information integrity and technical usability dimensions, in the technical solution of this application, the AI ​​intelligent image review server further performs an image quality assessment on the non-motorized vehicle violation image data to determine whether the image quality of the non-motorized vehicle violation image data meets preset requirements.

[0063] In practice, the server works collaboratively through a multi-layered evaluation model: the bottom layer uses traditional image algorithms to calculate global sharpness and local contrast, such as detecting whether there are continuous biometric contours in the facial area in a helmet-occluded scene; the middle layer uses a lightweight convolutional network to identify the continuity of the vehicle frame contour or the degree of occlusion of the license plate by rain streaks in rain- and fog-affected images; the top layer introduces a semantic segmentation module to verify the spatial integrity of illegal elements (such as faces, license plates, and non-motorized vehicle structures). This layered strategy can identify details lost due to underexposure in low-light scenes at night, or assess the resolvability of backlit areas in strong light and backlight scenes. In this way, a strong data foundation is provided for subsequent identity confidence calculations.

[0064] Specifically, in step S23, the AI ​​intelligent image review server determines whether the non-motorized vehicle violation image data is duplicate data. It should be understood that urban electronic eye systems often generate multiple images with high spatiotemporal overlap within seconds due to multiple cameras working together or the same violation continuously triggering cameras. For example, the same electric vehicle running a red light may be continuously captured by cameras in three directions at an intersection, or a vehicle may linger before the stop line, triggering multiple captures. If such redundant data enters the subsequent processing flow without filtering, it not only increases the computational load but also leads to multiple duplicate records for the same violator in the system, seriously interfering with the accurate statistics of violations and the uniqueness benchmark for identity comparison. Therefore, in the technical solution of this application, the AI ​​intelligent image review server determines whether the non-motorized vehicle violation image data is duplicate data. That is, intelligent deduplication of violation events is achieved through dynamic feature comparison.

[0065] In practice, the server employs a multimodal feature fusion strategy: First, it extracts spatiotemporal trajectory fingerprints from images (such as violation timestamps, geographic coordinates, and non-motorized vehicle movement vectors) to establish a basic spatiotemporal collision model. Second, it extracts visual fingerprints of the violating entity (such as helmet texture, vehicle structure, and local features of cyclist clothing) through a lightweight convolutional network, combining this with environmental semantic information (such as traffic light status and surrounding building outlines) to generate composite feature codes. Finally, it uses high-dimensional feature similarity calculation and adaptive time window threshold settings (e.g., setting consecutive images within 10 seconds of the same coordinate point as suspicious duplicates) to achieve correlation determination across camera data. This process effectively distinguishes between multiple triggers caused by brief stops and truly independent violations. This not only improves system computational efficiency but also ensures the uniqueness and statistical accuracy of violation database records, providing deduplicated, clean data input for the multi-source confidence model.

[0066] Taking the scheme in this application as an example, the system captures three violation records in a short period of time: East-facing camera at 09:00:03: The electric vehicle's left-turn red light illuminates as it crosses the stop line; the license plate number is clearly visible. South-facing camera at 09:00:05: The electric vehicle enters the center of the intersection; the rider's blue overalls are clearly visible. West-facing camera at 09:00:08: The electric vehicle is about to leave the intersection; the reflective strip on the rear rack is damaged. After receiving the data, the AI ​​review server initiates a deduplication process: First, the system extracts the spatiotemporal fingerprints of the three records: the timestamps are all within a 5-second interval of 09:00:03-09:08; the GPS coordinates are all matched to the same intersection center point (error < 3 meters); motion vector analysis shows that the electric vehicle is continuously traveling from east to west at a speed of 15 km / h. Based on the preset "10 seconds / same coordinate point" collision rule, the three data points were marked as a suspicious duplicate set. Then, the server simultaneously launched a lightweight convolutional network analysis: extracting the characteristics of the illegal subject from the three images respectively: the electric vehicle's front fender has the same L-shaped scratch (local feature similarity 98.7%); the rider's cuffs all have yellow reflective strips (texture matching 96.2%); the knot on the left side of the rear rack has a unique bow shape (structural consistency 99.1%); the outline of the "Golden Eagle Supermarket" neon sign in the background of the three images is completely consistent, and the traffic light state sequence conforms to the real switching logic of "red for left turn from east → red for straight traffic from north to south → red for left turn from west"; subsequently, the server encoded the above features into a 384-dimensional composite vector and calculated the pairwise similarity: if the similarity between pairs exceeds the 90% threshold, the system determines it as a continuous capture of the same illegal event. Based on the above results, the server automatically retains the first complete license plate image captured by the east-facing camera as the master record, marks the south-facing and west-facing data as redundant copies, and compresses the original three records into a single violation event and stores it in the database.

[0067] Specifically, in step S3, in response to the image review result being approved, multi-source data fusion and comparison are performed based on the non-motorized vehicle violation image data to obtain multi-source basic comparison result data. It should be understood that single-dimensional identity recognition methods are insufficient to address the complex challenges of actual law enforcement. When the violator wears a helmet, severely obscuring facial features, relying solely on image comparison is highly prone to failure; and for cyclists violating across regions, the lack of local registration records means that relying solely on a basic database query will lead to missed identifications. Furthermore, although the images after review have filtered out low-quality data, the identity clues they contain remain highly uncertain, necessitating cross-verification of multi-dimensional evidence chains to overcome the bottleneck of missing local information. Therefore, in the technical solution of this application, in response to the image review result being approved, multi-source data fusion and comparison are performed based on the non-motorized vehicle violation image data to obtain multi-source basic comparison result data.

[0068] Specifically, the system first analyzes identifiable elements in the violation images (such as the outline of eyebrows and eyes not completely obscured by the helmet, and local features of the vehicle) as initial seed information, simultaneously triggering multi-path verification: In the spatial dimension, it captures the violation location patterns of repeat offenders through high-frequency point correlation analysis (e.g., the same rider repeatedly running red lights at specific intersections); in the temporal dimension, it identifies continuous violation characteristics by combining historical violation records with real-time trajectory overlap calculations; in the asset dimension, it verifies non-motorized vehicle registration information to lock the identity of the person associated with the vehicle; and at the same time, it performs multi-party comparisons with the basic database to ensure the authority of the original data. This dynamic fusion of multi-dimensional data makes identity recognition no longer limited by the lack of single-point information, providing a highly robust multi-dimensional evidence base for subsequent confidence calculations.

[0069] In specific implementation, such as Figure 4 As shown, in response to the review result being approved, multi-source data fusion and comparison are performed based on the non-motorized vehicle violation image data to obtain multi-source basic comparison result data. This includes: comparing candidate identity documents with the real-name database based on the non-motorized vehicle violation image data to obtain intermediate data for image comparison and judgment; performing correlation analysis between candidate high-frequency locations and violation locations based on the non-motorized vehicle violation image data to obtain intermediate data for violation location correlation analysis; calculating candidate trajectory overlap based on the non-motorized vehicle violation image data to obtain intermediate data for candidate trajectory overlap; confirming the candidate's basic database multi-party comparison results based on the non-motorized vehicle violation image data to obtain intermediate data for basic database multi-party comparison confirmation; verifying candidate non-motorized vehicle registration information based on the non-motorized vehicle violation image data to obtain intermediate data for registration information verification; and querying candidate historical violation records based on the non-motorized vehicle violation image data to obtain intermediate data for historical violation record query.

[0070] Taking the solution of this application as an example, the image comparison of the candidate's real name database with the non-motorized vehicle violation image data is performed to obtain intermediate data for image comparison and judgment, including the following example process:

[0071] First, the driver's facial image is extracted from the non-motorized vehicle violation image data. In a specific example, the non-motorized vehicle violation image data is input into a face detection network based on the YOLO model to obtain the driver's facial image. Next, the driver's image is input into a third-party face comparison engine (e.g., Hikvision engine or Yitu engine) to obtain intermediate data for image comparison and judgment. This intermediate data includes the violation number, candidate name, ID number, similarity score, registration photo, and a large image of the violation capture. Specifically, the third-party face comparison engine includes a candidate real-name database, which is a large database of personal information, typically the permanent / temporary resident population database of the relevant management system (referred to as the "permanent / temporary database"). This database stores citizens' identity information, most importantly the associated standard ID registration photo (or document photo). It is worth noting that the intermediate data for image comparison and judgment is not the final judgment, but a structured intermediate result set. This result set summarizes and organizes candidate information from different face comparison engines.

[0072] To illustrate with an example, let's simulate a scenario: Wang was caught on camera running a red light at an intersection while riding an electric bike.

[0073] Step 1: Illegal data collection and facial recognition.

[0074] Time: 9:30 AM, June 9, 2024;

[0075] Location: Intersection of Renmin Road and Zhongshan Road in a certain city;

[0076] Incident: A person drove a blue electric vehicle and ran a red light.

[0077] System Operation: The non-motorized vehicle violation capture device at the intersection automatically starts, taking a set of high-definition photos (large image and small image) and a short video.

[0078] Data Generation: The system generates a unique violation number for this violation: VJ2024******00001. The AI-powered intelligent image review server performs a preliminary analysis of the image and confirms it as a valid non-motorized vehicle violation. The system successfully extracts the driver's facial image from the high-resolution photo and names it face_VJ2024******00001.jpg.

[0079] Step 2: Image comparison of the person and the identification document.

[0080] System Operation: The system sends the extracted face image face_VJ2024******00001.jpg to two integrated third-party face comparison engines—"Hik Engine" and "Yitu Engine".

[0081] The comparison process included: Hikvision Engine compared the data with the "Temporary Resident Real-Name Database" and returned the three candidates with the highest scores: Wang Moumou, ID number: 3301...001X, similarity: 92.5%; Li Moumou, ID number: 3301...002Y, similarity: 87.1%; Zhang Moumou, ID number: 3301...003Z, similarity: 85.4%.

[0082] Yitu Engine also performed a comparison and returned the three candidates with the highest scores: Wang Moumou, ID number: 3301...001X, similarity: 93.1%; Liu Moumou, ID number: 3301...004W, similarity: 88.0%; Li Moumou, ID number: 3301...002Y, similarity: 86.8%.

[0083] Step 3: Generate intermediate data for image comparison and analysis.

[0084] System Operation: The system receives and integrates the comparison results from the two engines. For the same violation number VJ2024******00001, all candidates who appeared (Wang Moumou, Li Moumou, Zhang Moumou, Liu Moumou) will be recorded. For the same person appearing in both result sets (such as Wang Moumou and Li Moumou), the system usually adopts the highest similarity score, or performs a weighted average (in this example, the highest value is taken).

[0085] Final output (written to the intermediate data table): Based on the data structure defined in the document, the system generates structured "intermediate data" records as shown in Table 1 below:

[0086] Table 1

[0087]

[0088] In this embodiment of the application, the intermediate data for the correlation analysis of violation points is obtained by performing a correlation analysis between high-frequency locations of candidates and violation locations based on the non-motorized vehicle violation image data, including the following example process:

[0089] First, the violation number and violation location number associated with the non-motorized vehicle violation image data are extracted; then, a candidate list is extracted from the intermediate data of the image comparison and analysis, which contains all candidates associated with this violation and their identity information (such as name and ID number).

[0090] Next, the system queries the historical vehicle capture data or personnel trajectory data for each candidate in the candidate list. In other words, the system needs to access the historical vehicle capture data or personnel trajectory database. Then, based on the historical vehicle capture data or personnel trajectory data for each candidate, the system determines the high-frequency activity points for each candidate to obtain a list of high-frequency activity points for each candidate. Specifically, for each person in the candidate list, the system will backtrack and query all their capture records within a past period (e.g., the past month), and then the system will count the number of captures at each point. If a certain point captures a candidate five or more times, that point is defined as a high-frequency activity point for that candidate.

[0091] Then, it is determined whether the violation location number exists in a candidate's high-frequency activity location list, and intermediate data for the violation location association analysis is generated based on the determination result. This intermediate data includes the violation number, candidate name, ID number, registration photo, high-frequency location number, high-frequency location name, and the number of times the location was captured. In other words, the system matches the location where the violation occurred with the high-frequency activity location list generated for each candidate in the previous step. If the violation location exists in a candidate's high-frequency location list, it indicates a strong spatial correlation between the candidate and the violation. If the violation location is not in a candidate's high-frequency location list, it indicates a weak spatial correlation between the candidate and the violation.

[0092] It's worth noting that the intermediate data from the illegal location correlation analysis specifically records and extracts information on candidates who violate the law in places the user frequently visits. This data clearly identifies which candidate, which high-frequency location, and how frequently the violation occurred (specifically, the number of times the image was captured), providing a crucial factor for subsequent confidence model calculations.

[0093] In one example, after completing the comparison of the identity images, the system identified Wang Moumou, Liu Moumou, Li Moumou, and Zhang Moumou as candidates.

[0094] Scenario: The system begins to perform correlation analysis between high-frequency locations and illegal locations for these four candidates.

[0095] Step 1: Determine the input information.

[0096] Current illegal data:

[0097] Violation ID: VJ2024******00001;

[0098] Location number of violation: P101 (intersection of Renmin Road and Zhongshan Road).

[0099] Candidate List:

[0100] Wang Moumou (ID number: 3301...001X);

[0101] Liu Moumou (ID number: 3301...004W);

[0102] Li Moumou (ID number: 3301...002Y);

[0103] Zhang Moumou (ID number: 3301...003Z).

[0104] Step 2: Core Processing - Calculate and correlate high-frequency points.

[0105] The system will process each candidate in turn:

[0106] Analysis of candidate [Wang Moumou]:

[0107] The system retrieved Wang's traffic violation records for the past month. It found that Wang lives nearby and passes through this intersection daily on her way to and from get off work. Statistics: Wang was captured 12 times at intersection P101 (Renmin Road and Zhongshan Road); and 20 times at intersection P102 (below her company building). Judgment: Because 12 ≥ 5 and 20 ≥ 5, the system identifies Wang's high-frequency violation locations as [P101, P102]. Correlation: The current violation location P101 exists in Wang's high-frequency violation location list [P101, P102]. [Successful Match]

[0108] Analyzing candidate [Liu Moumou]: The system retrieved Liu Moumou's traffic violation records for the past month. Findings: She lives on the other side of the city and just happened to be passing through this area today. Statistics: Liu Moumou was captured once at P101 (intersection of Renmin Road and Zhongshan Road); she was captured 25 times at P205 (entrance to her residential compound). Judgment: Because 1 < 5, P101 is not a high-frequency location for Liu Moumou. Her high-frequency location list is [P205]. Correlation: The current violation location P101 does not exist in Liu Moumou's high-frequency location list [P205]. [Matching failed];

[0109] Analyzing candidates [Li Moumou] and [Zhang Moumou]: The system performed the same analysis and found that both candidates had fewer than 5 snapshots at point P101, and their respective high-frequency activity points did not include P101. [Matching failed].

[0110] Step 3: Generate intermediate data for high-frequency location association.

[0111] System Operation: Based on the above analysis, only Wang meets the condition that "the illegal location is his high-frequency location". Therefore, the system will generate a record according to the data structure defined in the document and write it into the "Candidate High-Frequency Location Association" intermediate data table.

[0112] The final output is shown in Table 2 below (written into the intermediate data table):

[0113] Table 2

[0114]

[0115] In this embodiment of the application, the overlap degree of candidate trajectories is calculated based on the non-motorized vehicle violation image data to obtain intermediate data of candidate trajectory overlap, including the following example process:

[0116] It's understandable that a person's identity is determined not only by their appearance but also by their behavioral patterns (i.e., their spatiotemporal trajectory). If a candidate's recent activity trajectory is highly similar to the activity trajectory of a face captured at a crime scene, then it's highly likely they are the same person. This step goes a step further than high-frequency location correlation; it no longer looks at individual locations but at the similarity of a path (trajectory) composed of a series of locations.

[0117] In the example process, the system first extracts the driver's facial image from the non-motorized vehicle violation image data. Using this driver's facial image as the target, image search technology is used to find capture records that match the driver's facial image. Based on these capture records, the system generates the violator's activity trajectory. In other words, the system uses the face in the current violation photo as the target and employs image search technology to search a historical trajectory database from the past two weeks to find all capture records that match the violator's face. This results in a list composed of capture location numbers, representing the activity trajectory of this unknown violator over the past two weeks.

[0118] Similarly, for each candidate in the candidate list, image search technology is used to generate the activity trajectory of each candidate. It is worth mentioning that the standard registration photo of each candidate, such as an ID card photo, is used in the image search process. That is, the system performs an image search for each person in the candidate list. This time, the search target is the candidate's standard registration photo (ID card photo) to find all the snapshot records that match the candidate's registration photo, thereby generating a list of the activity trajectory of each candidate over the past half month.

[0119] Finally, the overlap rate between the offender's activity trajectory and the activity trajectories of each candidate is calculated to obtain intermediate data on candidate trajectory overlap. In a specific example, the overlap percentage = (number of common points in the two trajectories / total number of points in the offender's activity trajectory) * 100%. It should be understood that the intermediate data on candidate trajectory overlap includes the violation number, candidate name, ID number, capture overlap percentage, and ranking, where the ranking is based on the order of the capture overlap percentage.

[0120] In one example, the system has identified candidates such as Wang Moumou and Liu Moumou and completed high-frequency location analysis. Now, we move on to a more advanced calculation of trajectory overlap.

[0121] Step 1: Determine the input and generate the "violation capture trajectory".

[0122] Input: Violation ID: VJ2024******00001; Violator's face: The face extracted from the violation photo face_VJ...123.jpg.

[0123] Candidates: Wang Moumou, Liu Moumou, Li Moumou, Zhang Moumou and their registration photos.

[0124] System Operation: The system uses the "violation's face" to search the database of captured images from the past 15 days.

[0125] Generating "Violation Capture Trajectory": The system discovered that this unidentified face appeared at 4 different locations in the past 15 days. Violation Capture Trajectory (A) = {P101, P102, P105, P301} (4 points in total).

[0126] Step 2: Generate trajectories for candidates and calculate overlap.

[0127] The system will process candidates one by one:

[0128] Analysis of [Wang Moumou]: The system uses Wang Moumou's registration photo to search the capture database for the past 15 days. It generates Wang Moumou's trajectory (B): Since she lives and works in this area, her trajectories are very similar. For example, the trajectory is {P101, P102, P105, P110, P112}. It calculates the overlap: Trajectory A and Trajectory B share the following common points: {P101, P102, P105}, totaling 3. The overlap percentage = (3 / 4) * 100% = 75%.

[0129] Analysis of [Liu Moumou]: The system uses Liu Moumou's registration photo to search the database of snapshots taken in the past 15 days. It generates Liu Moumou's trajectory (C): She only passed through this location on the day of the violation; all other activities took place on the other side of the city. For example, the trajectory is {P101, P205, P210}. Overlap calculation: The common point between trajectories A and C is {P101}, with a count of 1. The overlap percentage is (1 / 4) * 100% = 25%.

[0130] The same analysis was performed on Li and Zhang, assuming that their trajectories overlapped even less, at 10% and 0% respectively.

[0131] Step 3: Generate intermediate data on trajectory overlap.

[0132] System Operation: The system now has the trajectory overlap scores for all candidates and can rank them accordingly. The final output is shown in Table 3 below (written into the "Candidate Trajectory Overlap" intermediate data table):

[0133] Table 3

[0134]

[0135] In this embodiment of the application, the intermediate data for confirmation of the candidate database through multi-party comparison based on the non-motorized vehicle violation image data is as follows:

[0136] It is understandable that in the field of facial recognition, no matter how accurate a single algorithm is, it may still make misjudgments or biases. By introducing multiple independent recognition engines from different vendors and requiring them to produce similar high scores for the same target, the reliability and confidence of identity recognition can be greatly improved.

[0137] In the specific process, the driver's face image is first extracted from the non-motorized vehicle violation image data. In a specific example, the non-motorized vehicle violation image data is input into a face detection network based on the YOLO model to obtain the driver's face image.

[0138] Next, the driver's face image is input into multiple face comparison engines (e.g., Yitu Face Comparison Engine and Zhili Face Comparison Engine) to obtain multiple candidate lists and their similarity scores. For example, when the multiple face comparison engines are Yitu Face Comparison Engine and Zhili Face Comparison Engine, the output results are a first candidate list and its similarity score and a second candidate list and its similarity score.

[0139] The system filters candidates from multiple candidate lists and their similarity scores, selecting those deemed highly similar by all or multiple face comparison engines simultaneously. For example, similarity scores must simultaneously exceed 90% to obtain intermediate data confirmed by multi-party comparisons in the base database. This intermediate data includes violation numbers, candidate names, ID numbers, registration photos, Zhili face comparison similarity, and Yitu face comparison similarity. In other words, its core processing aims to filter out candidates deemed highly similar by all (or multiple) face comparison engines. This is a very strict filtering rule, as described in the documentation, containing two core conditions (dual AND logic): Co-occurrence: A candidate must appear in both the Yitu engine's and Zhili engine's results. If a candidate is only identified by one engine, he / she will be directly filtered out. And, Dual High-Score Threshold: For candidates meeting the "co-occurrence" condition, the system further checks their similarity scores. The candidate must have a similarity score greater than 90% according to Yitu Engine and also greater than 90% according to Zhili Engine. If either score is not met, the candidate will be filtered out.

[0140] It should be understood that the intermediate data confirmed by multiple comparisons in the base library is a high-confidence candidate list generated after rigorous screening. Typically, this list is very short, and may even be empty. This intermediate data represents an identification result with almost no controversy. The marked candidates are suspects jointly identified by multiple authoritative algorithms, and all have extremely high scores. In the final confidence model, although the weight score of this item itself may not be high, it is a strong confirmation signal, effectively supporting and verifying the validity of the high-weighted single-item highest similarity result in the first step. Simultaneously, the intermediate data confirmed by multiple comparisons in the base library records the verified candidates in detail and lists their similarity scores in each engine, facilitating traceability and display.

[0141] In one example, we'll use the case of Wang Moumou running a red light. The system has already obtained a preliminary candidate list including Wang Moumou, Liu Moumou, and others through comparisons across various engines. Now, we'll begin confirming the results of the multi-engine comparison.

[0142] Step 1: Organize the independent comparison results of each engine.

[0143] Violation information: Violation ID: VJ2024******00001; and the violator's face: face_VJ...123.jpg.

[0144] Input - Yitu Engine's comparison results (top few):

[0145] Wang Moumou, ID number: 3301...001X, similarity: 93.1%;

[0146] Liu Moumou, ID number: 3301...004W, similarity: 88.0%.

[0147] The comparison results of the input-ZhiLi engine (top few):

[0148] Wang Moumou, ID number: 3301...001X, similarity: 92.5%;

[0149] Mr. Zhou, ID number: 4403...987V, similarity: 90.5%;

[0150] Li Moumou, ID number: 3301...002Y, similarity: 87.1%.

[0151] Step 2: Core Processing - Applying Filtering Rules.

[0152] The system will review each candidate appearing in any list individually:

[0153] Review of [Wang Moumou]: Does he / she appear in both Yitu and Zhili's results? Yes. Yitu similarity 93.1% > 90%? Yes. Zhili similarity 92.5% > 90%? Yes. Conclusion: All conditions are met, pass the screening.

[0154] Review of [Liu Moumou]: Does it appear in both Yitu and Zhili results? No (appears only in Yitu results). Conclusion: Co-occurrence is not satisfied; it has been filtered.

[0155] Review of [Mr. Zhou]: Does it appear in both Yitu and Zhili results? No (appears only in Zhili results). Conclusion: Co-occurrence is not satisfied; it has been filtered out.

[0156] Review of [Li Moumou]: Does it appear in both Yitu and Zhili results? No (appears only in Zhili results). Conclusion: Co-occurrence is not satisfied; it has been filtered.

[0157] Step 3: Generate intermediate data that has been verified through multi-party comparison of the base library.

[0158] System Operation: After rigorous "consultation" and screening, only Wang Moumou passed the verification. The system will generate a record for her. The final output is shown in Table 4 below (written to the candidate base database, and an intermediate data table is also output):

[0159] Table 4

[0160]

[0161] In this embodiment of the application, the intermediate data for verifying the registration information of candidate non-motorized vehicles based on the non-motorized vehicle violation image data is as follows:

[0162] It is understandable that if a violation is caused by a non-motorized vehicle, then a candidate who owns a registered non-motorized vehicle is naturally more likely to commit the crime than a candidate who does not own a registered vehicle.

[0163] In a specific process example, the candidate list is first extracted from the intermediate data of the image comparison and analysis. That is, the candidate list is first obtained from the initial identity verification step, and each person in the list is associated with their own standard ID photo (registration photo).

[0164] Next, image search technology is used to determine whether each candidate in the candidate list is a known non-motorized vehicle owner. Specifically, the standard identity registration photo of each candidate is used as the query image and input into the target database (the target database is the facial images of all vehicle owners in the non-motorized vehicle registration database) to determine whether each candidate is a known non-motorized vehicle owner through image search technology, for example, if the similarity exceeds 90%. That is, the system will traverse the candidate list, using each candidate's registration photo to perform a 1:N search in the vast database of registered faces. If the system finds a registered face whose similarity to the candidate's registration photo exceeds a preset threshold, then the corresponding candidate is determined to be a known non-motorized vehicle owner.

[0165] Based on the query results, intermediate data for verifying the registration information is generated. The intermediate data for verifying the registration information includes the violation number, candidate name, ID number, registration photo, non-motor vehicle registration database face URL, and face comparison similarity.

[0166] In one example, continuing with Wang's case, after she and several other candidates were listed, the system initiated a verification of the non-motorized vehicle registration information.

[0167] Step 1: Determine the input and verification background.

[0168] Input: Violation ID: VJ2024******00001; Candidate list: [Wang Moumou, Liu Moumou, Li Moumou, Zhang Moumou] and their respective standard ID registration photo URLs.

[0169] Background information: Ms. Wang purchased a Yadea electric scooter in 2023 and registered it at the community service center. During registration, staff used specialized equipment to take her facial photo, which was then stored in the "Non-Motor Vehicle Registration and Filing Database" along with her identity information. Ms. Liu purchased her electric scooter several years ago and had not been registered locally. Mr. Li and Mr. Zhang also do not have any registered non-motor vehicles under their names.

[0170] Step 2: Core Processing - Individual Verification. The system begins iterating through the candidates and performing image comparison:

[0171] Verification of [Wang Moumou]: The system retrieved Wang Moumou's standard ID registration photo. Using this photo as the query image, a 1:N image search was performed in the "Non-Motor Vehicle Registration and Filing Database." Result: The system successfully matched Wang Moumou's facial photo taken during the 2023 registration process in the database. Calculation showed a facial similarity of 95.8% between the ID registration photo and the filing photo. Judgment: 95.8% > 90% (preset threshold), verification passed. Wang Moumou is confirmed as the registered owner of the non-motor vehicle.

[0172] Verification of [Liu Moumou]: The system retrieved Liu Moumou's standard ID registration photo and compared it with the "Non-motorized Vehicle Registration and Filing Database." Result: The system failed to find any registered face with a similarity exceeding 90% to Liu Moumou's registration photo. Judgment: Verification failed.

[0173] Verification of [Li Moumou] and [Zhang Moumou]: The system comparison also failed the verification.

[0174] Step 3: Generate intermediate data for registration information verification. System operation: In this verification, only Wang Moumou meets the criteria. Therefore, the system will generate a record for her and store it in the corresponding intermediate table. The final output is shown in Table 5 below (the intermediate data table containing the candidate's non-motorized vehicle registration information):

[0175] Table 5

[0176]

[0177] In this embodiment of the application, the process of querying candidate historical violation records based on the non-motorized vehicle violation image data to obtain intermediate data for the historical violation record query includes the following example steps:

[0178] It's understandable that if a candidate has a prior record of non-motorized vehicle violations that have been resolved, they are more likely to commit similar violations again compared to someone without such a record. This is a method of assessing the likelihood of current behavior from a behavioral history perspective.

[0179] During its execution, the system first extracts a candidate list from the intermediate data of the image comparison and analysis. Next, it uses image search technology to determine whether any candidate in the list has a past record of non-motorized vehicle violations that have been processed. In specific implementation, each candidate's standard ID photo is used as the query image and input into a target database (the target database consists of captured facial images associated with all processed violation records in the historical violation database) to obtain query results. These results are used to determine whether any candidate has a past record of non-motorized vehicle violations that have been processed. In a specific example, the criterion is whether the similarity exceeds 90%. That is, the system uses each candidate's ID photo to perform a 1:N search in the historical violation facial database. If a candidate's ID photo successfully matches a captured facial image in the historical violation database, and the similarity exceeds a preset threshold, then it is confirmed that the corresponding candidate has a past record of non-motorized vehicle violations that have been processed.

[0180] Based on the query results, intermediate data for querying historical violation records is generated. The intermediate data for querying historical violation records includes violation number, candidate name, ID number, URL of the registration photo, URL of the face of the historical violation, large image of the historical violation, and face comparison similarity.

[0181] In a specific example, let's consider the scenario of Wang Moumou and several other candidates. The system now needs to check whether these candidates have a history of non-motorized vehicle violations.

[0182] Step 1: Determine the input and background.

[0183] Input: Violation number (this time): VJ2024******00001 (Note that this table only records the context of this violation); Candidate list: [Wang Moumou, Liu Moumou, Li Moumou, Zhang Moumou] and their respective standard registration photo URLs.

[0184] Background Information (Simulated): Wang: She was penalized once in 2022 for riding an electric bike without a helmet (the penalty has been processed). Her face was also captured by the system at the time. Liu, Li, and Zhang: These individuals have no prior records of non-motorized vehicle violations in the historical database.

[0185] Step 2: Core Processing - Searching the historical database one by one, the system performs a search for each candidate.

[0186] Query for

Wang Moumou

[0187] Query for [Liu Moumou]: The system compared Liu Moumou's registration photo with the historical database. Result: No facial similarity exceeding 90% was found in the database for any person with a historical record of illegal activity. Judgment: No match found.

[0188] Searching for "Li Moumou" and "Zhang Moumou": similarly, no historical violations were found.

[0189] Step 3: Generate intermediate data for historical violation record queries. System operation: Only Wang Moumou matches this rule. The system will generate a record for her. The final output is shown in Table 6 below (written into the intermediate data table of candidates' historically processed violations):

[0190] Table 6

[0191]

[0192] Specifically, in step S4, the multi-source basic comparison result data is input into the multi-source data fusion identity DNA confidence model to obtain personnel identity confidence score data. It should be understood that although the six types of intermediate data generated through multi-source fusion comparison provide multi-dimensional identity clues, their evidentiary value differs fundamentally and their interrelationships are complex. When faced with scenarios such as large-area helmet occlusion leading to fluctuations in image matching credibility, or missing registration information for cross-regional violators, relying solely on human experience or rule engines cannot accurately quantify the true contribution weight of each dimension of evidence in the overall identity determination. This is especially true when multi-source data presents contradictions (e.g., high trajectory overlap but low image matching), making it difficult for traditional methods to construct a scientific decision-making system. Therefore, in the technical solution of this application, the multi-source basic comparison result data is input into the multi-source data fusion identity DNA confidence model to obtain personnel identity confidence score data.

[0193] In specific implementation, the intermediate data from image comparison and judgment, the intermediate data from illegal location association analysis, the intermediate data from candidate trajectory overlap, the intermediate data from multi-party comparison and confirmation in the basic database, the intermediate data from registration information verification, and the intermediate data from historical violation record query are first input into the multi-source data fusion identity DNA confidence model to obtain the first to sixth personnel identity confidence score items. During this process, the model first performs feature distillation and evidence strength transformation on the six types of intermediate data: image comparison and judgment data extracts the credibility factor of matching features through a convolutional attention mechanism; illegal location association data uses a spatiotemporal pattern recognition algorithm to quantify the statistical significance of behavioral inertia; trajectory overlap data uses dynamic time warping technology to evaluate the continuity strength of movement paths; and registration verification and historical record data verify the correlation strength between assets and behaviors based on information completeness. Further, the weighted sum of the first to sixth personnel identity confidence score items is calculated to obtain the personnel identity confidence score data. Here, the confidence scores can be linearly aggregated using the optimal weight configuration obtained through pre-training (image comparison as the dominant weight, and the remaining auxiliary weights) to form the final identity DNA confidence spectrum. In calculating the weighted sum of the confidence scores for the first to sixth individuals to obtain the individual identity confidence scores, the weights of these individual confidence score components are 88%, 2%, 2%, 2%, 3%, and 3%, respectively. This dynamic balancing mechanism of evidence weighting enables the system to maintain a high level of confidence in cross-regional illegal cases, even with missing information in the basic database, based on spatiotemporal behavioral patterns. The final generated individual identity confidence scores overcome the limitations of traditional binary judgment, forming a continuous spectral deterministic metric, providing a quantitative decision-making benchmark for subsequent differentiated processing strategies.

[0194] In the embodiments of this application, it should be understood that the intermediate data for image comparison and judgment includes: How similar does this person look? (basic score); intermediate data for illegal location correlation analysis includes: Did he / she commit the crime in the "territory"? (spatial correlation); intermediate data for trajectory overlap includes: Does his / her recent activity path match that of the violator? (behavioral pattern); intermediate data for multi-party comparison and confirmation includes: Do all experts (algorithms) agree that it is him / her? (cross-validation); intermediate data for filing information verification includes: Does he / she have the "tools for committing the crime"? (contextual correlation); and intermediate data for historical violation query includes: Has he / she done this before? (behavioral inertia).

[0195] In the process of inputting the multi-source data fusion identity DNA confidence model to obtain the identity confidence scores of the first to sixth persons, if a candidate appears in the corresponding intermediate data (including intermediate data for image comparison and judgment, intermediate data for illegal location association analysis, intermediate data for candidate trajectory overlap, intermediate data for multi-party comparison and confirmation in the basic database, intermediate data for filing information verification, and intermediate data for historical violation record query), the model adds the full weight score of this rule to him / her. For example, if there is a historical violation, add 2 points; or if a candidate appears in the intermediate data for image comparison and judgment with a similarity of 95%, then his / her corresponding score is 95 points.

[0196] In a specific example, continuing with the case of Ms. Wang, suppose the system is scoring her and another candidate, Ms. Liu. Step 1: Summarize all the intermediate data evidence. For the illegal event VJ2024******00001, the system has compiled the evidence for the two main candidates, as shown in Table 7 below:

[0197] Table 7

[0198]

[0199] Step 2: The model calculates based on the weight table.

[0200] The scoring engine reads the preset weight configuration table as shown in Table 8 below:

[0201] Table 8

[0202]

[0203] The model begins calculating sub-scores for each candidate:

[0204] For [Wang Moumou], the following scores are calculated: Image comparison score = 0.931 * 88 = 81.928 points; Basic library output score = 2 points (because a record exists); Filing information score = 2 points (because a record exists); Historical violation score = 2 points (because a record exists); Trajectory overlap score = 3 points (because a record exists); High-frequency location score = 3 points (because a record exists).

[0205] For [Liu Moumou], the following scores are calculated: Image comparison score = 0.880 * 88 = 77.44 points; Basic library output score = 0 points (because no record exists); Filing information score = 0 points (because no record exists); Historical violation score = 0 points (because no record exists); Trajectory overlap score = 3 points (because her trajectory also matched, although the overlap was low); High-frequency location score = 0 points (because no record exists).

[0206] Step 3: Output the confidence score data for the first to sixth individuals. The system ultimately generates the following structured data. Record 1: Violation ID: VJ2024******00001; Candidate: Wang Moumou; Item_data:{ "First Confidence Score": 81.928}{ "Second Confidence Score": 2}{ "Third Confidence Score": 2}{ "Fourth Confidence Score": 2}{ "Fifth Confidence Score": 3}{ "Sixth Confidence Score": 3}. Record 2: Violation ID: VJ2024******00001; Candidate: Liu Moumou; Item_data:{ "First Confidence Score": 77.44}{ "Second Confidence Score": 0}{ "Third Confidence Score": 0}{ "Fourth Confidence Score": 0}{ "Fifth Confidence Score": 3}{ "Sixth Confidence Score": 0}.

[0207] Specifically, S5 determines a differentiated processing strategy based on the personnel identity confidence score data. It is understandable that non-motorized vehicle violation identification faces a contradiction between the need for massive data processing and limited law enforcement resources. This is especially true when helmet obstruction leads to inconsistent image quality, and cross-regional behavior causes information gaps. If a uniform processing procedure is applied to all violation records, it will cause delays in efficient case review (e.g., clear faces running red lights) or incorrect entry of inefficient cases into the database (e.g., mismatches in rain or fog). Traditional methods do not distinguish between differences in evidence strength, making it difficult to cope with the daily throughput pressure of tens of thousands of law enforcement data points, and are prone to systemic misjudgment risks due to the accumulation of low-quality data. Therefore, to construct a confidence-driven dynamic resource allocation mechanism and achieve the optimal balance between law enforcement accuracy and efficiency through a hierarchical decision-making model, in the technical solution of this application, in response to the personnel identity confidence score data belonging to high-confidence personnel, the differentiated processing strategy is to directly process the data into the database without manual review; in response to the personnel identity confidence score data belonging to medium-confidence personnel, the differentiated processing strategy is to process the data into the database after manual review and confirmation; in response to the personnel identity confidence score data belonging to low-confidence personnel, the differentiated processing strategy is to directly invalidate the data. In this way, a two-dimensional optimization of law enforcement resources and the credibility of identification results is achieved.

[0208] In summary, the confidence-based non-motorized vehicle and pedestrian identification method according to the embodiments of this application is explained. It achieves dynamic initial screening of illegal images through an AI intelligent image review server, using a deep neural network to deconstruct the essential differences between image quality defects and missing target features, accurately filtering invalid data. Subsequently, multi-source data fusion and comparison are initiated for approved images, innovatively integrating multi-source data into an identity DNA confidence model. This model dynamically adjusts the contribution of each data source through an adaptive weight allocation mechanism, ultimately outputting a continuously quantified identity confidence score. Thus, by deeply coupling the original image quality assessment with the dynamic fusion of multi-source data, a complete confidence chain from data cleaning to hierarchical judgment is constructed, eliminating the consumption of computing resources by invalid data. This enables the verification capability of violator identity confidence scores based on a pedestrian and non-motorized vehicle confidence model, effectively reducing the workload of police officers reviewing violations, increasing the rate of public education and punishment, and promoting the reform of the duty mechanism.

[0209] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A confidence-based non-motorized vehicle pedestrian identification method, characterized in that, The method comprises the following steps: S1, acquiring non-motor vehicle illegal image data collected by a camera; S2, inputting the non-motor vehicle illegal image data into an AI intelligent film review server to obtain a film review result; S3, in response to the film review result being passed, performing multi-source data fusion comparison based on the non-motor vehicle illegal image data to obtain multi-source basic comparison result data; S4, inputting the multi-source basic comparison result data into a multi-source data fusion identity DNA confidence model to obtain personnel identity confidence score data; S5, determining a differential processing strategy based on the personnel identity confidence score data; S2 comprises the following steps: S21, the AI intelligent film review server performs illegal behavior identification and illegal object identification on the non-motor vehicle illegal image data to determine whether there is an illegal identification error; S21 comprises the following steps: performing illegal object visual feature extraction based on convolutional coding on the non-motor vehicle illegal image data to obtain an illegal object visual feature map; performing feature unitization on the illegal object visual feature map along the channel dimension to obtain a sequence distribution of channel unit illegal object visual feature vectors; based on the channel saliency of each channel unit illegal object visual feature vector in the sequence distribution of channel unit illegal object visual feature vectors, performing sparse processing and feature shape remodeling on the sequence distribution of channel unit illegal object visual feature vectors to obtain an illegal object visual feature sparse coding map, including: inputting the sequence distribution of channel unit illegal object visual feature vectors into a channel feature interrelation encoder based on a Transformer architecture to obtain a sequence distribution of channel unit context-related illegal object visual feature coding vectors; calculating the channel saliency of each channel unit context-related illegal object visual feature coding vector in the sequence distribution of channel unit context-related illegal object visual feature coding vectors; based on the channel saliency of each channel unit context-related illegal object visual feature coding vector, performing sparse processing on the sequence distribution of channel unit illegal object visual feature vectors to obtain a sparse sequence distribution of channel unit illegal object visual feature vectors; performing feature shape remodeling on the sparse sequence distribution of channel unit illegal object visual feature vectors to obtain an illegal object visual feature sparse coding map; based on the illegal object visual feature sparse coding map, determining whether the illegal object is a non-motor vehicle to obtain an illegal object identification result.

2. The confidence-based non-motorized vehicle pedestrian identification method of claim 1, wherein, inputting the non-motor vehicle illegal image data into an AI intelligent film review server to obtain a film review result further comprises the following steps: the AI intelligent film review server performs image quality evaluation on the non-motor vehicle illegal image data to determine whether the image quality of the non-motor vehicle illegal image data meets the preset requirements; the AI intelligent film review server determines whether the non-motor vehicle illegal image data is repeated data. 3.The confidence-based non-motorized vehicle pedestrian identification method of claim 1, wherein, in response to the film review result being passed, performing multi-source data fusion comparison based on the non-motor vehicle illegal image data to obtain multi-source basic comparison result data, comprising: Based on the non-motor vehicle illegal image data, candidate person real name library image comparison is performed to obtain intermediate data of image comparison research; Based on the non-motor vehicle illegal image data, candidate person high-frequency point and illegal point association analysis is performed to obtain intermediate data of illegal point association analysis; Based on the non-motor vehicle illegal image data, candidate person trajectory coincidence degree calculation is performed to obtain intermediate data of candidate person trajectory coincidence; Based on the non-motor vehicle illegal image data, candidate person basic library multi-party comparison result confirmation is performed to obtain intermediate data of basic library multi-party comparison confirmation; Based on the non-motor vehicle illegal image data, candidate person non-motor vehicle record information verification is performed to obtain intermediate data of record information verification; Based on the non-motor vehicle illegal image data, candidate person historical illegal record query is performed to obtain intermediate data of historical illegal record query.

4. The confidence-based non-motorized vehicle pedestrian identification method of claim 3, wherein, The multi-source basic comparison result data is input into a multi-source data fusion identity DNA confidence model to obtain personnel identity confidence score data, including: The intermediate data of image comparison research, the intermediate data of illegal point association analysis, the intermediate data of candidate person trajectory coincidence, the intermediate data of basic library multi-party comparison confirmation, the intermediate data of record information verification, and the intermediate data of historical illegal record query are input into the multi-source data fusion identity DNA confidence model to obtain first to sixth personnel identity confidence score sub-item data; The weighted sum of the first to sixth personnel identity confidence score sub-item data is calculated to obtain the personnel identity confidence score data.

5. The confidence-based non-motorized vehicle pedestrian identification method of claim 4, wherein, Based on the personnel identity confidence score data, a differentiated processing strategy is determined, including: In response to the personnel identity confidence score data belonging to high confidence score personnel, the differentiated processing strategy is direct storage processing without manual review; In response to the personnel identity confidence score data belonging to medium confidence score personnel, the differentiated processing strategy is manual review confirmation before storage; In response to the personnel identity confidence score data belonging to low confidence score personnel, the differentiated processing strategy is direct invalidation processing.

6. The confidence-based non-motorized vehicle pedestrian identification method of claim 1, wherein, Based on the illegal object visual feature sparse coding graph, it is determined whether the illegal object is a non-motor vehicle to obtain an illegal object recognition result, including: The illegal object visual feature sparse coding graph is input into a classifier-based illegal object recognizer to obtain a classification result, which is used to indicate whether the illegal object is a non-motor vehicle.

Images