Non-motor vehicle overspeed identification method and device based on image sequence

Abstract

The application provides a non-motor vehicle overspeed identification method and device based on an image sequence, relates to the technical field of image analysis, and comprises the following steps: acquiring image sequence data collected by a non-motor vehicle in a low-light curve channel; performing feature parameter extraction based on the image sequence data, and performing mapping processing based on the extracted feature parameters to obtain an embedded trajectory point set of the non-motor vehicle in the curve channel; performing direction-sensitive tensor graph construction and multi-scale turning perception graph modeling processing to obtain a multi-dimensional graph structure; performing angular graph attention mechanism and convolution trajectory transformation network processing to obtain a speed estimation vector sequence; performing acceleration filtering and compensation processing to obtain a speed sequence from which turning inertia error is removed; and performing curve overspeed behavior analysis to obtain a determination result of whether the non-motor vehicle has a curve overspeed behavior. The application realizes fine reconstruction of a non-motor vehicle trajectory, inertia error filtering and speed behavior analysis by using a multi-stage algorithm combination.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image analysis, in particular to a non-motor vehicle overspeed identification method and device based on image sequences. BACKGROUND

[0002] Currently, non-motor vehicles (such as electric bicycles, shared bicycles, etc.) are widely used in urban roads and have become an important part of urban travel. With the increase in usage frequency, the traffic hazards caused by the overspeed behavior of non-motor vehicles in complex traffic environments have gradually become prominent, especially in low-light underground passages, closed three-dimensional passages, and narrow spaces with curves, overspeed riding can easily cause serious traffic accidents such as collisions and falls. The existing non-motor vehicle speed detection methods mostly rely on single-frame image target detection, manual setting of speed thresholds, or electronic speed limit devices. These methods have two significant problems: first, it is difficult to accurately estimate the true speed trajectory of non-motor vehicles in continuous image sequences, especially in turning, occlusion, or image blur situations; second, without a pre-set speed upper limit or road markings, it is difficult to effectively distinguish between normal riding and risky overspeed behavior. In addition, traditional motion modeling methods cannot accurately express the curvature change, turning trend, and inertial interference of non-motor vehicles on curved road sections, resulting in low speed anomaly recognition accuracy, false positives and false negatives, and difficulty in meeting the actual needs of fine traffic behavior management in complex scenarios.

[0003] Therefore, there is an urgent need for a non-motor vehicle overspeed identification method and device based on image sequences to solve the above technical problems. SUMMARY

[0004] The purpose of the present application is to provide a non-motor vehicle overspeed identification method and device based on image sequences to improve the above problems. In order to achieve the above purpose, the technical solutions adopted by the present application are as follows:

[0005] In a first aspect, the present application provides a non-motor vehicle overspeed identification method based on image sequences, comprising:

[0006] obtaining image sequence data collected by a non-motor vehicle in a low-light curved passage;

[0007] extracting non-motor vehicle boundary features and passage structure parameters based on the image sequence data, and performing mapping processing based on the extracted features and passage structure parameters to obtain an embedded trajectory point set of the non-motor vehicle in the curved passage;

[0008] performing direction-sensitive tensor graph construction and multi-scale turning perception graph modeling processing on the embedded trajectory point set to obtain a multi-dimensional graph structure containing time dimension, spatial direction, and turning curvature information;

[0009] The multi-dimensional graph structure is processed through an angular graph attention mechanism and a convolution trajectory transformation network to obtain a speed estimation vector sequence of the non-motor vehicle;

[0010] The speed estimation vector sequence is subjected to acceleration filtering and compensation processing to obtain a speed sequence with removal of turning inertia error;

[0011] The speed sequence is subjected to non-motor vehicle curve overspeed behavior analysis to obtain a determination result of whether the non-motor vehicle has curve overspeed behavior.

[0012] In a second aspect, the present application further provides a non-motor vehicle overspeed identification device based on an image sequence, comprising:

[0013] An acquisition unit is configured to acquire image sequence data collected by a non-motor vehicle in a low-light curve channel;

[0014] A mapping unit is configured to extract non-motor vehicle boundary features and channel structure parameters based on the image sequence data, and perform mapping processing based on the extracted features and channel structure parameters to obtain an embedded trajectory point set of the non-motor vehicle in the curve channel;

[0015] A modeling unit is configured to perform direction-sensitive tensor graph construction and multi-scale turning perception graph modeling processing on the embedded trajectory point set to obtain a multi-dimensional graph structure containing time dimension, spatial direction and turning curvature information;

[0016] A processing unit is configured to process the multi-dimensional graph structure through an angular graph attention mechanism and a convolution trajectory transformation network to obtain a speed estimation vector sequence of the non-motor vehicle;

[0017] A removal unit is configured to perform acceleration filtering and compensation processing on the speed estimation vector sequence to obtain a speed sequence with removal of turning inertia error;

[0018] An analysis unit is configured to perform non-motor vehicle curve overspeed behavior analysis on the speed sequence to obtain a determination result of whether the non-motor vehicle has curve overspeed behavior.

[0019] The present application has the following beneficial effects:

[0020] The present application is aimed at low light conditions, closed underground passageway and irregular road scene with continuous curve structure, and proposes a complete technical process of fusion direction sensitive graph structure modeling, curvature guided manifold mapping, angular attention mechanism, trajectory convolution estimation and speed anomaly discrimination fusion algorithm, which obtains image sequence data and constructs a multi-dimensional graph model with spatial direction, time continuity and structural curvature information, realizes fine reconstruction of non-motor vehicle trajectory, inertia error filtering and speed behavior analysis by using multi-stage algorithm combination, and finally realizes adaptive identification of speeding behavior in curved road section without presetting speed threshold or artificial limiting, which significantly improves the robustness, precision and application range of the system in complex environment, and overcomes the core defects of the prior art such as dependence on rules, one-sided recognition and poor adaptability.

[0021] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent from the description, or can be learned by practice of the present application as hereinafter described. The objects and other advantages of the present application will be realized and attained by means of the structures particularly pointed out in the written description and claims hereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments, and it should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0023] Figure 1 The flow chart of the non-motor vehicle speeding identification method based on image sequence described in the embodiments of the present application;

[0024] Figure 2 The structure diagram of the non-motor vehicle speeding identification device based on image sequence described in the embodiments of the present application.

[0025] In the figure: 701, acquisition unit; 702, mapping unit; 703, modeling unit; 704, processing unit; 705, removal unit; 706, analysis unit. DETAILED DESCRIPTION

[0026] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0027] It should be noted that similar reference numerals and letters refer to similar items throughout the drawings, and therefore, once an item is defined in one drawing, it need not be further defined and explained in subsequent drawings. Meanwhile, in the description of the present application, the terms "first", "second", and the like are only used to distinguish descriptions, and cannot be understood as indicating or implying relative importance.

[0028] Embodiment 1

[0029] The embodiment provides a non-motor vehicle overspeed identification method based on an image sequence.

[0030] Referring to Figure 1 , the method includes steps S1, S2, S3, S4, S5 and S6.

[0031] Step S1, acquiring image sequence data collected by a non-motor vehicle in a low-light curve channel;

[0032] It can be understood that the image sequence data in this step includes not only image frames at consecutive time stamps, but also non-motor vehicle edge feature extraction signals under weak light conditions, channel space structure modeling parameters (such as bend curvature, channel width, inclination, wall material), and light reflection models in the environment (such as ground and wall reflection coefficients, light arrangement influenced reflection trajectories), image frames at consecutive time stamps are obtained from image acquisition devices installed inside the enclosed underground channel (such as low-illumination cameras, infrared imaging equipment or wide-dynamic-range cameras), non-motor vehicle edge feature extraction signals under weak light conditions refer to signals obtained by introducing weak light image enhancement algorithms (such as low-light Retinex enhancement, global contrast enhancement, infrared contour compensation, etc.) during image acquisition, and the geometric modeling parameters of the channel structure and the light reflection model are usually obtained in advance through construction design drawings or on-site scanning modeling systems (such as laser SLAM or three-dimensional point cloud modeling), and then modeling is performed through these data. This part provides the original data basis with time continuity, structural clarity and spatial positioning availability for subsequent boundary feature extraction and trajectory embedding modeling, and at the same time, it constructs a perception input entrance that can be compatible with low light and complex terrain, thereby significantly improving the adaptability and stability of the system in non-ideal visual environment.

[0033] Step S2, based on the image sequence data, non-motor vehicle boundary feature and channel structure parameter extraction is performed, and mapping processing is performed based on the extracted features and channel structure parameters to obtain a set of embedded trajectory points of the non-motor vehicle in the curved channel;

[0034] It can be understood that this step extracts core features with geometric and motion significance from enhanced image frames and structural parameters provided in the image acquisition stage to establish an embeddable motion trajectory relationship between the non-motor vehicle and the curved channel. This step not only converts the distributed fuzzy and directionally unstable edge information in the image frames into a spatially continuous and structurally reliable motion trajectory point set, but also introduces physical modeling of the channel structure in the entire process, so that the trajectory point set has a measurable geometric basis in subsequent speed modeling, solving the existing problems of large motion path reconstruction error, serious boundary distortion and non-physical pixel trajectory in complex curved structures. In this step, step S2 includes step S21, step S22, step S23 and step S24.

[0035] Step S21, boundary tensor response extraction processing is performed on the image sequence data, a preliminary non-motor vehicle boundary response map is obtained by constructing a multi-frequency guided edge tensor response model of the non-motor vehicle candidate region in each image frame of the image sequence data;

[0036] It can be understood that this step introduces a multi-frequency guided edge tensor response model, decomposes the image signal in different frequency domains, suppresses the reflection area and shadow interference in the low-frequency part, and strengthens the boundary contour response in the high-frequency part, especially the structural change characteristics in the motion direction. Among them, this step uses wavelet transform or Gabor filter group to construct the multi-frequency tensor atlas of the image, and on this basis, it guides the edge response combined with the motion candidate area (the area where the non-motor vehicle may exist is pre-screened by inter-frame difference or optical flow estimation) to construct a boundary response tensor with direction sensitivity and frequency domain stability.

[0037] This processing not only targets the static pixel intensity in the image frame, but also combines the time evolution characteristics in the image sequence, uses the time sequence consistency verification mechanism to further filter the boundary results, eliminates false edges caused by transient bright spots or background structure mutations, and retains the structural contour evolution trend of non-motor vehicles in consecutive frames. The final output of the preliminary non-motor vehicle boundary response map not only contains spatial edge distribution information, but also includes direction features and confidence labels, providing a high-quality structure foundation for the next step of boundary curvature modeling and channel geometry mapping.

[0038] Step S22, performing local boundary curvature estimation processing on the preliminary non-motor vehicle boundary response map, fitting the non-motor vehicle boundary points in multiple scales, and introducing a second-order guiding term in the strong turning area to suppress image noise interference to obtain a boundary curvature point set;

[0039] It can be understood that this step first performs multi-scale direction fitting processing on the boundary point set detected in the boundary response map. In the neighborhood of each candidate boundary point, a direction fitting window is constructed, and polynomial curve fitting or local Bezier curve model is used to perform curvature approximation in multiple scales (such as 3, 5, 7, 11 pixel neighborhoods), to extract the first-order direction derivative vector, and combine the local boundary trend to perform initial curvature estimation. The multi-scale strategy is particularly important in this scenario: because in the low-light curve, the non-motor vehicle may have local static (causing the edge to be too smooth) or rapid turning (causing the edge to be deformed) phenomena in one frame, different scales can respectively cope with the contradictory needs of noise suppression and boundary preservation.

[0040] In order to further enhance the geometric stability of the fitting result, this step introduces a second-order guiding term in the strong turning area (i.e. the boundary segment with sharp curvature change and sharp direction fluctuation), modifies the first-order fitting direction drift by constructing the curvature acceleration term between the boundary points, and suppresses the local abnormal jump caused by reflection, shadow or texture interference in the image. This guiding mechanism essentially adds a regularization term to the curvature estimation function, so that it has boundary following property in the turning section without causing oscillation, thereby outputting a boundary curvature point set with smooth direction, real curvature and continuous boundary, and taking it as the non-motor vehicle boundary feature.

[0041] Step S23, the boundary curvature point set is guided by channel structure perspective geometry modeling processing, through the use of channel structure parameters of image sequence data to map the trajectory in the image frame from the pixel domain to the curvature space in multiple stages, to obtain the estimated motion path sequence of the non-motor vehicle;

[0042] It can be understood that this step is through the camera perspective parameters (including focal length, inclination, projection center) in the image sequence data and the structure parameters (such as curve radius, channel width, wall orientation) of the channel, wherein the channel structure parameters are obtained by uploading the channel construction data collected by the staff, constructing the initial affine mapping matrix from the image pixel point to the world coordinate system, and performing basic space projection on the boundary curvature point set. This mapping is used to roughly eliminate the projection distortion, so that the spatial distribution of the boundary points matches the actual channel structure. Then, for the channel area with strong curve curvature, a nonlinear curvature inversion model is introduced: the curvature information of each boundary point in the image frame is fitted with the "geometric curve profile function" in the channel parameters to construct the mapping relationship from the image coordinates to the curvature space, so that each trajectory point has a real spatial explanation of the local bending trend.

[0043] It can be understood that the geometric curve profile function in this step is as follows:

[0044] ;

[0045] Wherein, represents the geometric curvature value of the path at the curve length parameter , represents the arc length parameter along the motion path of the non-motor vehicle, represents the spatial curvature of the path at .

[0046] And this step also introduces a timing consistency constraint, traces the spatial transformation trend of the boundary curvature points in the image sequence, constructs a dynamic alignment function of the trajectory points between different frames by fusing the movement law of the same point trajectory between multiple frames, so that the motion trajectory forms a continuous and derivable path form in the physical space. These mapping outputs finally form a set of estimated motion path sequences arranged in time sequence, with spatial consistency and physical structure basis.

[0047] Step S24, the motion path sequence is processed by manifold embedding, to obtain the embedded trajectory point set of the non-motor vehicle in the curve channel.

[0048] It can be understood that this step is first based on the time sequence adjacency relationship between the path points, the local curvature change rate and the spatial distribution density to construct a similarity graph structure of the trajectory points, the local neighborhood of each point not only considers the time sequence continuity, but also introduces a spatial turning trend as a weight adjustment factor to depict the embedded differences between different motion modes such as 'accelerated turning' and'slow sliding' in the path. On the basis of this graph structure, this step uses the Riemann manifold embedding algorithm to project the nonlinearly distributed path points in the high-dimensional space into the embedding space, so that the global path topology is maintained while the expression ability of the local motion structure (such as sharp bends and speed jump points) is strengthened.

[0049] At the same time, in order to avoid local overfitting or graph rupture in the embedding process, this step introduces an isometric mapping mechanism to ensure that the physical distribution consistency between the path points is still maintained in the embedding space, that is, the high-density section such as the entrance of the curve is still concentrated, and is not artificially stretched after embedding, thereby improving the interpretability of the embedded path and the stability of the subsequent graph construction.

[0050] Step S3, direction-sensitive tensor graph construction and multi-scale turning perception graph modeling processing are performed on the embedded trajectory point set to obtain a multi-dimensional graph structure containing time dimension, spatial direction and turning curvature information;

[0051] It can be understood that this step is based on the trajectory point set obtained after manifold embedding, and further constructs a multi-dimensional tensor graph structure that integrates time series, spatial direction and geometric curvature information. This graph structure not only serves as a carrier for the spatial connection relationship of the trajectory, but also is a core data support form for subsequent speed estimation and behavior recognition. This step significantly improves the adaptability of the model to behavior pattern variability, and can effectively distinguish between low-speed large-bend and high-speed small-bend behaviors. In this step, step S3 includes step S31, step S32, step S33 and step S34.

[0052] Step S31, the embedded trajectory point set is subjected to local direction encoding processing, wherein the motion direction between adjacent trajectory points is represented as a directed edge in the tensor to obtain an initial direction tensor graph;

[0053] It can be understood that the system of this step first traverses each pair of time-adjacent trajectory points in the embedded trajectory point set, calculates the spatial position difference vector by the Euclidean distance calculation method, extracts the unit direction vector (for example, constructs the direction angle based on the change in three-dimensional space), and then encodes the direction information as edge weight or edge attribute into the trajectory graph structure, so that each edge in the graph not only connects two trajectory points, but also explicitly carries the information of'moving from which point to which direction'. This process actually establishes a structure expression layer of a direction field covering the trajectory graph, and each edge can be regarded as a directional motion element path.

[0054] Different from traditional undirected trajectory graphs or simple time-series point sets, the direction-encoding tensor graph can capture important behavior characteristics such as local direction mutation, continuous turning trend, and motion inertia trend in subsequent processing, and is particularly suitable for identifying complex dynamic patterns of turning areas. In addition, the system introduces a tensor structure (i.e., multi-channel encoding) while constructing the graph, and each tensor channel can represent different dimensions of direction correlation, such as the main direction, normal vector, and historical direction mean, further enhancing the representation ability of the graph structure to local behavior characteristics.

[0055] Step S32, performing spatial direction consistency optimization processing on the initial direction tensor graph, wherein a direction average field in a spatial neighborhood is introduced in the initial direction tensor graph, and the direction average field is used to smooth irregular turning change areas to obtain an optimized tensor graph structure.

[0056] It can be understood that this step first defines a spatial neighborhood window with a fixed radius for each trajectory point in the embedding space, extracts all adjacent trajectory points connected to the point through a graph edge in the window, and collects the direction vectors thereof. Then, a weighted average direction of the neighborhood direction vectors is calculated, wherein the weight can be adjusted according to the time difference of the trajectory points (a higher weight is given to a point closer to the time point), the curvature similarity (a penalty is given to a direction mutation), or the local density (to avoid misleading in sparse areas), to form a direction estimation field representing the “expected motion trend” of the point.

[0057] Next, this step uses the average direction field to adjust the edge attributes of the initial tensor graph that have abnormal direction mutations. For edges with a direction deviating too much from the average direction of the neighborhood, the system replaces the direction vector of the edge with the average direction in the direction field, or uses a regularization interpolation method to smooth the adjustment. Finally, the connection relationship between all trajectory points is optimized in terms of direction characteristics while keeping the time sequence unchanged, forming an optimized tensor graph structure with continuous edge attributes, high direction consistency, and strong robustness to local disturbances.

[0058] Step S33, performing multi-scale turning response modeling processing on the optimized tensor graph structure, extracting feature responses corresponding to preset turning patterns by layering modeling of turning gradient information in different scales in the tensor domain, and obtaining a turning perception graph structure.

[0059] It can be understood that this step first defines multiple direction perception scales (such as short scale 3 nodes, medium scale 5-7 nodes, and long scale 11 or more nodes) on the optimized tensor graph structure. At each scale, the direction change sequence (i.e. direction gradient) of the continuous point pairs in the trajectory is extracted, and indexes such as turning angle change rate, cumulative angle increment, and local direction change density are calculated to construct the direction gradient tensor graph layer at this scale. Each scale layer is essentially a direction response graph with a specific field of view range, and its edge attribute describes the turning trend and change intensity of the trajectory within the scale range.

[0060] Subsequently, this step models the multi-scale graph layers in the tensor domain: by constructing a cross-scale response stack structure, the feature responses of each scale are combined according to scale weight and structure stability to form a multi-scale turning response graph with hierarchical perception ability. In order to extract behavior features with recognition significance, this step performs pattern matching on each scale layer, performs structure similarity matching and response extraction according to the prior-defined typical turning mode template (such as sharp turning, continuous fine-tuning, and composite turning), and finally integrates the recognized turning structure mode back into the corresponding trajectory segment in the tensor graph to generate a turning perception graph structure with structure label and multi-scale direction response features.

[0061] Step S34, performing multi-dimensional graph structure fusion processing on the turning perception graph structure, by jointly fusing trajectory timestamps, position coordinates, direction vectors, and local curvature encoding into high-order graph node features, to obtain a multi-dimensional graph structure containing time dimension, spatial direction, and turning curvature information.

[0062] It can be understood that this step first expands the feature dimension of each trajectory node (i.e. an embedded trajectory point) in the turning perception graph. Among them, the following four types of information are introduced for each node and jointly encoded:

[0063] Timestamp feature: each trajectory point corresponds to the frame time in the image sequence, indicating its relative time position in the entire trajectory evolution process. This feature is used to maintain the sequential logic of the trajectory points in time, and is the time basis for identifying dynamic behaviors such as speed change and acceleration trend.

[0064] Position coordinate feature: the coordinate value of the trajectory point in the embedding space, which preserves its geometric position in the low-dimensional curvature space, supporting spatial consistency analysis and path geometry alignment.

[0065] Direction vector feature: represents the motion direction of the non-motor vehicle at this point, derived from the edge attribute of the direction tensor graph in the previous step, and is the core information for analyzing behavior inertia, turning trend, and momentum continuity.

[0066] Local curvature encoding: acquired from the boundary curvature estimation stage, used to quantify the degree of bending of the path at this point, especially for identifying "sharp turning points", "coasting sections" and other spatial areas with specific dynamic risks.

[0067] The above four types of information are integrated into a multi-dimensional vector through a high-order node feature splicing mechanism. The system constructs a multi-channel tensor representation of the graph nodes based on this, i.e., each node contains its temporal position, directionality, geometric shape, and spatial state. In order to maintain the dimensional consistency and expression stability between features, normalization and scale adjustment are also performed on each channel (e.g., standardizing time and unitizing direction vectors), so that different dimensional features can participate in learning together in the graph neural network without dominant dimensional bias problems.

[0068] Finally, the high-dimensional feature tensors of all nodes constitute the node attribute set of the entire multi-dimensional graph structure, and the edge attributes inherit the aforementioned direction weight, turning amplitude, etc. information, forming a space-time-direction-curvature four-coupled graph structure.

[0069] Step S4, processing the multi-dimensional graph structure through the angular graph attention mechanism and the convolutional trajectory transformation network to obtain a speed estimation vector sequence of the non-motor vehicle;

[0070] It can be understood that this step jointly uses the angular graph attention mechanism and the convolutional trajectory transformation network to deeply model the trajectory graph structure containing multi-dimensional information such as time, spatial direction, and curvature, which is constructed in the previous step. In order to derive the speed estimation vector sequence of the non-motor vehicle from complex trajectory behaviors, this step utilizes the structure perception ability of the graph neural network and the direction selection ability of the angle attention mechanism. For diversified, nonlinear, and frequently locally disturbed trajectory behaviors in the curve channel, this step introduces a direction perception graph attention mechanism and a curvature adaptive convolutional trajectory network. The system can extract a speed estimation result with high precision, low volatility, and strong robustness to abnormal structure changes in the graph structure, especially in the curve channel, which can avoid speed misestimation or pseudo-acceleration phenomena, providing a high-trust speed basis for subsequent acceleration filtering and abnormal behavior recognition. In this step, step S4 includes step S41, step S42, step S43, and step S44.

[0071] Step S41, initializing the attention weight of the angular partial derivative guide of the multi-dimensional graph structure, obtaining the initial weight matrix of the graph attention by taking the angular offset of each node in its directional neighborhood in the multi-dimensional graph structure as the kernel weight of the attention calculation;

[0072] It can be understood that for any one trajectory node in the figure, first the direction vector of each adjacent node is extracted in its direction neighborhood (i.e. the set of nodes pointed to by the directed edges connecting the node), and the angle between the direction vector of the trajectory node and the current direction vector is calculated to form an angular offset. The system then constructs an attention kernel function based on an angular partial derivative function (such as a cosine function or a Gaussian direction kernel) to map the angular offset to the corresponding attention strength. In this step, nodes with close angles have higher direction consistency and should be given higher weights in information fusion; while neighbor nodes with large direction mutations may correspond to turning jitter, slip or discontinuous trajectory segments, and their attention weights will be weakened or eliminated.

[0073] In order to further improve the stability of the direction weight calculation, a normalization mechanism (such as softmax or L1 normalization) of the direction weight is introduced in this step to ensure that the sum of the adjacent weights of each node is 1, forming a learnable attention distribution. This initial weight matrix serves as the basis input for the graph attention mechanism, which will be used as the basis for dynamic regulation of edge weights in the subsequent graph neural network propagation and speed estimation process.

[0074] Step S42, performing node neighborhood feature aggregation processing on the initial weight matrix of the graph attention to fuse direction, time and curvature information in the variable neighborhood of the node, forming an aggregated feature tensor to obtain a set of node state vectors;

[0075] It can be understood that this step uses the feature data in the adjacent node set of each trajectory node in the graph to perform weighted aggregation. The features of each adjacent node include three core channels: direction vector, reflecting the current motion trend of the trajectory; timestamp, describing the time sequence evolution position of the trajectory point; and local curvature value, measuring the bending degree of the path at the node. The system weights these features by guiding the attention coefficients related to each adjacent node in the guide weight matrix, and performs the following weighted combination:

[0076] ;

[0077] wherein, is a nonlinear projection operation on the original feature vector of the adjacent node , including feature scaling, channel normalization and other modules, for unifying the representation scale of different physical quantities (direction as vector, time as scalar, curvature as scalar). Through this operation, this step generates a direction-time-structure coupled state vector at each node , which accurately reflects the geometric state, evolution trend and local path complexity of the current trajectory point in its motion context, represents the attention coefficient of node to adjacent node , and a set of adjacent nodes of the node, denotes an adjacent node a raw feature vector of the node, denotes a coupled node state vector.

[0078] In this step, the aggregation process adopts a variable neighborhood mechanism, that is, the neighborhood size of the node is not fixed, but is dynamically adjusted according to the trajectory density, direction continuity and structural consistency in the graph, allowing more detailed aggregation in dense path segments (adapt to high frequency direction changes), and expanding the aggregation field in sparse areas to enhance stability.

[0079] Step S43, the node state vector set is processed by trajectory convolution transformation, the sliding window convolution operation based on the direction sequence is performed on the order time axis of the trajectory, and the channel curvature is introduced as the convolution weight adjustment factor, to obtain a convolution feature map reflecting the dynamics of the trajectory time sequence;

[0080] It can be understood that in this step, all node state vectors in the trajectory graph are first reordered according to their time stamps, and a time-ordered state vector sequence is constructed. Then a fixed-size sliding window (such as length 5) is defined, and a one-dimensional convolution operation is performed on the sequence along the time axis. Unlike standard one-dimensional convolution, the convolution kernel of the system not only slides in the time dimension, but also combines the direction vector change trend of each window internal node for direction consistency weighting, that is, if the direction changes sharply in a window, the direct contribution of this segment of convolution output to the speed estimation is reduced, in order to avoid the misleading caused by sharp turns or path disturbances.

[0081] In addition, a channel curvature adjustment factor is introduced as an external control parameter of the convolution weight in this step. The specific method is: for each sliding window, the average curvature value of the nodes in the window is calculated as a geometric complexity index of the region; the system dynamically adjusts the convolution kernel weight response according to the curvature size - in high curvature (i.e. sharp turn) area, the perception ability of direction gradient change is enhanced, and in low curvature area, the filtering effect of speed stability is enhanced. This "structure-guided convolution" mechanism can be regarded as a kind of geometric modulation means for the trajectory convolution process, so that the feature map retains the path structure difference in convolution.

[0082] The finally output convolution feature map is a tensor structure in time sequence, which contains the comprehensive feature response of each trajectory point in its local time-direction neighborhood. The feature map not only encodes the motion trend of the non-motor vehicle at each time, but also reflects its historical evolution path and the motion change that will occur, providing a dynamic representation with context awareness ability for subsequent speed estimation.

[0083] Step S44, the convolution feature map is subjected to a velocity vector analysis process, and a first-order and second-order derivative analysis is performed on the node sequence in the convolution output domain to obtain a velocity estimation vector sequence of the non-motor vehicle.

[0084] It can be understood that this step first performs a first-order derivative analysis on the node state sequence in the convolution feature map, which is used to capture the speed change trend between consecutive time points. Since the feature vector of each node contains direction, position and time semantics, the system takes the time stamp as the main variable of the derivative, calculates the rate of change of each node feature between adjacent time points, and obtains the instantaneous speed estimation value. This process is equivalent to establishing a speed trend curve in the feature space, and the slope of the curve corresponds to the average speed of the object on the path, while the nonlinear component caused by the change in direction is also integrated, so that the estimation not only depends on the position difference, but also reflects the influence of trajectory shape change on the speed.

[0085] Secondly, this step further performs a second-order derivative analysis on the velocity sequence, that is, the rate of change of velocity with respect to time is derived to obtain the acceleration trend, which is used to identify inertial driving, rapid acceleration or deceleration behavior. In the curved channel, this high-order information is particularly important, because the non-motor vehicle often exhibits a short "sliding" or "speed delay drop" phenomenon during turning, and a first-order estimation alone may misjudge the true speed state. Through the second-order derivative analysis, this step can identify and suppress such speed fluctuations caused by inertial driving, thereby improving the physical accuracy and behavior consistency of the speed estimation.

[0086] Each speed value in the entire velocity vector sequence in this step has a clear time sequence position, structural source and direction coupling background, forming a group of high-confidence and behavior-perception speed estimation values.

[0087] Step S5, the velocity estimation vector sequence is subjected to acceleration filtering and compensation processing to obtain a speed sequence with removed turning inertia error;

[0088] It can be understood that this step effectively solves the problem of serious interference of speed estimation by direction fluctuation and inertial slip in the curved environment through curvature-aware acceleration filtering and trajectory structure-driven speed compensation, thereby improving the stability, reliability and structural consistency of the speed sequence, providing a high-quality and error-controllable input signal for the behavior recognition module, and significantly enhancing the application reliability of the system in actual traffic scenarios. In this step, step S5 includes step S51, step S52, step S53 and step S54.

[0089] Step S51, the velocity estimation vector sequence is subjected to local acceleration response extraction processing, the acceleration of each time node in the velocity estimation vector sequence is calculated, and the direction switching rate is used as a suppression factor to obtain an original acceleration vector sequence of the non-motor vehicle.

[0090] It can be understood that in this step, the system calculates the rate of change of velocity of each pair of adjacent velocity estimation points to obtain the preliminary instantaneous acceleration. This process not only considers the numerical change, but also synchronously extracts the trajectory direction vectors corresponding to the two time points to calculate the rate of change of the included angle, which is defined as the direction switching rate. This is one of the key innovations of this step: in the process of turning, U-turning or sliding, although the numerical change of the speed is significant, such a change is often caused by direction switching or inertial driving, rather than real acceleration behavior.

[0091] Therefore, the direction switching rate is introduced as a dynamic inhibition factor in the acceleration analysis in this step: if the acceleration is high in a certain period, but the direction switching rate is also high, it means that the acceleration is likely to be caused by trajectory redirection rather than dynamic driving, and its credibility should be reduced; on the contrary, in the area where the direction remains stable but the speed increases sharply, the acceleration is retained as an effective motion feature. This inhibition mechanism is realized through a threshold inhibition function, which essentially uses the direction stability as a weight function of the acceleration credibility to filter out false judgments caused by direction jumps.

[0092] Finally, the original acceleration vector sequence output by this step not only contains the acceleration response at all time nodes, but also contains the direction coupling background of each acceleration, so that the subsequent filtering and compensation operations can make judgments based on structural information.

[0093] Step S52, performing inertial disturbance identification processing on the original acceleration vector sequence, identifying a non-real speed fluctuation region dominated by a turning trajectory structure by setting a turning inertia detection interval in a high-curvature segment embedded in the trajectory, to obtain an inertial disturbance time period marking set;

[0094] It can be understood that in this step, all high-curvature segments in the embedded trajectory structure are first extracted, i.e., trajectory regions in which the curvature value of the continuous trajectory points is higher than a set threshold. These high-curvature regions usually correspond to sharp turns, continuous bends or rapid rotation structures, which are spatial regions where inertial disturbances are prone to occur. Subsequently, the system sets a turning inertia detection interval in each high-curvature segment, and the interval length can be dynamically adjusted according to the local time span of the trajectory, the curvature change rate and the direction switching strength. The interval not only includes the curvature peak position, but also extends a certain time window before and after, so as to cover the entire propagation stage of the inertial influence, in which the inertial disturbance is a lateral centripetal acceleration signal caused by the turning motion of the vehicle. It is a disturbance signal that will pollute the true readings of the accelerometer in the forward direction of the device, resulting in errors when calculating the speed. The purpose of this step is to mark these "unreliable" data segments to provide a basis for subsequent data processing.

[0095] Within these detection intervals, this step combines the original acceleration vector sequence, analyzes the acceleration abnormal amplitude, change trend and direction consistency, and identifies the "pseudo-wave segment" that does not conform to the power driving rule. For example: when the acceleration significantly deviates from the average value of the neighborhood, and the change direction is consistent with the path curvature direction, but there is no trend of continuous speed growth, this segment can be marked as inertial disturbance. Finally, a set of inertial disturbance period markers is generated, which contains all suspected non-real speed anomaly time periods caused by turning structures, and each marker is attached with its corresponding curvature interval and disturbance intensity score.

[0096] This step uses the geometric curvature distribution of the trajectory to perform structural constraint analysis on the speed fluctuation. The system can accurately identify the "non-dynamic speed disturbance" area caused by the superposition of nonlinear path morphology and motion inertia, and improve the discrimination ability between "real speed behavior" and "path-induced error".

[0097] Step S53, performing speed smoothing sequence reconstruction processing on the set of inertial disturbance period markers, obtaining a preliminary speed smoothing sequence of the non-motor vehicle by weighted estimation of the speed values in the marked period;

[0098] It can be understood that this step first defines the reconstruction range of each marked disturbance period, i.e. the speed sub-sequence containing the start and end boundaries of the period. Then, the speed values in the sub-sequence are weighted and estimated according to three dimensions, including:

[0099] Direction stability weighted estimation: evaluate the direction change rate corresponding to each speed point in the disturbance segment. The more stable the direction, the higher the weight. The direction mutation area is considered to be strongly disturbed by inertia, and should be given a lower weight to reduce its impact on the speed fitting curve.

[0100] Neighborhood speed trend consistency weighted estimation: by comparing with the speed trend of the front and rear non-disturbance areas, it is judged whether the disturbance segment deviates from the overall motion trend. Points consistent with the trend are kept with higher trust, and points deviating from the trend are given lower weight to achieve continuous repair of the trajectory speed.

[0101] Local curvature adjustment factor weighted estimation: areas with high curvature are more susceptible to disturbance. This step automatically reduces the speed response of high-curvature points based on the pre-set penalty factor of curvature value, and strengthens the dynamic adaptation to the change of path morphology.

[0102] The weights of the above three dimensions are set by presetting corresponding thresholds, and then the speed values in the disturbance segment are re-estimated by using a weighted sliding window average method or a first-order fitting interpolation method. In the reconstruction process, it is ensured that the speed curve is consistent in slope and coherent in trend at the beginning and end of the disturbance segment and the undisturbed area, avoiding the formation of new discontinuity or speed mutation in the reconstruction result. For some sliding segments or areas with large directional switching, Bezier smoothing or spline curve interpolation can also be used to enhance the naturalness of the transition.

[0103] Finally, the preliminary speed smoothing sequence output by this step not only retains the real speed change trend in the trajectory, but also eliminates the fluctuation abnormalities caused by path morphology or inertial error, providing a high-quality, low-noise speed basis for the next step of global speed curve re-calibration and behavior judgment.

[0104] Step S54, performing global trend calibration processing on the preliminary speed smoothing sequence, wherein the overall speed curve is re-fitted under the premise of ensuring energy conservation to obtain a speed sequence that removes the inertial error of the turn.

[0105] It can be understood that the preliminary speed smoothing sequence is first subjected to global trend analysis to evaluate the speed change trend in the entire trajectory range, especially for areas with large-scale changes such as turning, deceleration, and acceleration. This step introduces a physical constraint condition to ensure that the change in speed is consistent with the physical conservation of energy.

[0106] Then, this step globally fits the entire speed sequence to generate a smooth speed curve that conforms to the physical motion law through a fitting algorithm such as least squares method, Bezier curve fitting, or spline interpolation. During fitting, this step dynamically adjusts the fitting weight according to the speed change in the turning area, especially in high-curvature or sharp-turning areas, the system gives the fitting model a higher weight to ensure that the speed change in these areas is more stable.

[0107] This step also adds a local correction factor during fitting to specifically correct the inertial error that may occur during turning or deceleration, so that the speed change during turning can be reasonably corrected without affecting the speed expression of straight-line segments or smooth driving segments

[0108] Finally, a speed sequence that removes the inertial error of the turn is obtained, which can truly reflect the actual driving speed of the non-motor vehicle on the curve and other complex road segments, eliminating short-term speed fluctuations caused by inertial sliding or path turning, ensuring that the speed estimation value is more smooth and reliable.

[0109] Step S6, performing curve overspeed behavior analysis on the speed sequence to obtain a judgment result of whether the non-motor vehicle has curve overspeed behavior.

[0110] It can be understood that this step can accurately determine whether the non-motor vehicle has a curve overspeed behavior through comprehensive analysis of the speed sequence and the road structure, and can adapt to complex road environments in real time. Compared with the traditional overspeed detection method, this step not only considers the fixed speed threshold, but also combines the road geometric characteristics, the actual dynamic behavior of the non-motor vehicle, and the time sequence consistency analysis, thereby improving the accuracy and sensitivity of the overspeed detection. In particular, in complex curve, low light and other environments, the overspeed behavior can be effectively identified and misjudgment can be avoided. This provides a more accurate analysis tool for traffic monitoring, traffic safety management and intelligent transportation systems. In this step, step S6 includes step S61, step S62, step S63 and step S64.

[0111] Step S61, encoding the speed sequence, jointly encoding each speed value in the speed sequence and the spatial curvature and direction change rate of the corresponding embedded trajectory point as an embedded vector to obtain a speed representation set of the non-motor vehicle;

[0112] It can be understood that this step extracts the speed value of each trajectory point and the spatial curvature (i.e. the bending degree of the path at the point) corresponding to the point. The curvature reflects the nonlinearity of the path. For example, in a curve, the curvature of the path is large; while in a straight line section, the curvature is close to zero. By jointly encoding the speed value and the curvature value of the point, the speed representation of each node can not only reflect the instantaneous speed of the motion, but also describe the geometric characteristics of the point on the path. Nodes with high curvature values (such as sharp turns) will get more attention related to their speed, because the motion of such nodes usually has high risk.

[0113] In addition to the spatial curvature, this step also takes the direction change rate (i.e. the rate of change of the steering angle) of the node as an important dimension of encoding. The direction change rate can measure the degree of steering intensity of the non-motor vehicle within a certain time. For example, the direction change rate will be higher when making a large turn, while the change rate is close to zero when driving smoothly.

[0114] For each trajectory point, this step calculates the relationship between the direction change rate and the time interval and encodes it as one of the features of the point together with the speed. This feature helps the system understand whether the motion state of the non-motor vehicle is smooth, whether there is acceleration, deceleration or coasting behavior, especially in the curve turning area, the direction change rate can prompt the possible inertial effect or steering intensity in the motion process.

[0115] Finally, by jointly encoding the speed value, the curvature and the direction change rate, these information is integrated into a high-dimensional embedded vector. In this way, each trajectory point not only contains its instantaneous speed information, but also includes its spatial features and steering features, forming an embedded representation rich in multi-dimensional information.

[0116] Step S62: Perform category distribution learning processing on the speed representation set. Estimate the normal curve speed embedding distribution in the preset historical trajectory and calculate the degree of deviation of the current speed representation set under the distribution to obtain a density deviation score vector that reflects the degree of speed anomaly.

[0117] Understandably, this step first utilizes the normal speed sequences from historical trajectory data under similar curve environments to construct a speed embedding distribution model. This model includes speed values ​​and their corresponding embedding vectors under different curve types and speed ranges. Through statistical analysis of the characteristics of these normal trajectories, this step can learn the typical motion patterns of non-motorized vehicles in various curve scenarios and construct a speed embedding space distribution model, which is a probability density function (such as Gaussian distribution or multidimensional kernel density estimation). In this process, the invention will group historical trajectories according to different curve types, vehicle behavior patterns (such as fast turns, smooth driving, etc.), and environmental factors (such as road friction, speed limits, etc.) and learn the speed distribution characteristics of each group.

[0118] Once the current velocity representation set is obtained, this step compares these velocity embedding vectors with the normal velocity embedding distribution in the historical trajectory. Specifically, for each current velocity vector, its probability density value under the normal velocity distribution is calculated, i.e., the degree of matching between the current velocity and the historical distribution. If the current velocity is near the center of the distribution, it indicates normal behavior; if the current velocity deviates significantly from the center of the normal distribution, it suggests that the velocity value may indicate abnormal behavior.

[0119] The formula for calculating its probability density value under the normal velocity distribution is as follows:

[0120] ;

[0121] in, This represents the probability density value of the current velocity vector under the normal velocity distribution; Represents the current velocity vector. The dimension of the velocity vector. It is pi. The determinant of the covariance matrix of the velocity vectors represents the variance and correlation between features; The inverse of the covariance matrix is ​​used to weight the deviations of the velocity vector, reflecting the relative importance of different features. This is the mean vector of the normal velocity distribution, representing the average level of various velocity characteristics in the historical trajectory. It is the transpose symbol. represents the square of Mahalanobis distance, which measures the distance between the current speed and the normal speed distribution. The larger the value of this term, the greater the difference between the current speed and the historical normal behavior, and the smaller the probability density value.

[0122] Based on these matching degrees, the density deviation score is calculated, which is the normalized inverse of the probability density. The score represents the deviation of the current speed representation set from the historical normal speed distribution. A lower deviation score indicates that the current speed is close to the historical normal trajectory, and a higher deviation indicates an abnormal speed.

[0123] By calculating the deviation of all nodes, a density deviation score vector is obtained, which reflects the degree of speed anomaly at each time node. Nodes with higher scores usually indicate speed anomalies, which may be caused by excessive speed, inertia sliding or incorrect estimation due to path structure.

[0124] In step S63, the density deviation score vector is classified into behavior state labels. By classifying the high deviation value region in the density deviation score vector, a set of local behavior state labels for the preset time period is obtained.

[0125] It can be understood that a score threshold is set according to the aforementioned density deviation score to distinguish between normal and abnormal behavior. For example, when the density deviation score exceeds a certain preset threshold, it indicates that the speed in this period may have abnormal behavior, such as excessive acceleration or deviation from the normal trajectory. The system classifies the deviation values into multiple levels, such as: low deviation value (score close to 0): indicates that the current speed is within the normal range and there is no abnormal behavior; medium deviation value: indicates that there is some abnormal fluctuation, which may be caused by inertia sliding or other short-term disturbances; high deviation value: indicates that the speed changes significantly, which may be excessive speed, sudden acceleration or other sudden behavior.

[0126] According to the classification results, the system generates local behavior state labels for each time period. For example: in the low deviation interval (normal state), the label is "normal driving"; in the medium deviation interval (suspicious behavior), the label is "abnormal fluctuation" or "potential sliding"; in the high deviation interval (abnormal behavior), the label is "excessive speed" or "acceleration anomaly".

[0127] Finally, the system matches each time period's local behavior state label with the corresponding trajectory point and time period according to the results of the behavior classification. Each behavior label not only represents a trajectory segment, but also associates the specific time period and behavior intensity of the segment, so that subsequent analysis can be accurate to the time period level, providing rich information for subsequent decision-making.

[0128] Step S64: Perform global trajectory determination processing on the set of local behavior state labels. By determining whether each local state label in the set of local behavior state labels is connected or continuous in the trajectory graph structure, the determination result of whether the non-motorized vehicle has speeding behavior on curves is obtained.

[0129] Understandably, this step first evaluates the temporal continuity of each label in the local behavior state label set. For example, if the "speeding" label appears multiple times in a trajectory and these labels are consecutive in the time series, it indicates that the behavior is likely a continuous speeding behavior, rather than a single speed anomaly. This step also calculates the persistence of the local behavior labels within a time window, that is, whether the state labels exhibit similar patterns in multiple adjacent time periods. For example, on a curve, if the "speeding" label persists for multiple time frames, the behavior is judged to be continuous speeding; if the "speeding" label appears only occasionally, it may only be a local anomaly.

[0130] Next, this step associates the local behavior state labels with nodes in the trajectory graph structure. Through connectivity analysis, it can be determined whether different local behavior labels form a continuous chain of behaviors in the trajectory graph. For example, if a speeding behavior label and a subsequent speed anomaly label are connected in the trajectory graph and exhibit temporal continuity, then these labels may reflect a complete speeding behavior trajectory. Specifically, this involves checking whether the various local behavior states in the label set can be connected by edges (representing the connection relationship between trajectory points) or continue continuously in time within the graph structure, using path connectivity in the graph to determine whether there is a continuous speeding behavior.

[0131] Finally, based on connectivity and persistence analysis, this step combines the criteria for determining cornering speeding to comprehensively assess whether cornering speeding behavior exists. If, on a certain corner segment of the trajectory, there are prolonged and continuous "speeding" behavior labels, and these labels form a continuous chain of behavior through connected paths in the trajectory map, the system will determine that the trajectory contains cornering speeding behavior. This step will also assess whether the speeding behavior meets the speeding standard based on its duration and intensity. If the speeding behavior occurs instantaneously on the corner, it may only be an isolated incident, but if it is a continuous high-deviation behavior, it is judged as risky speeding behavior.

[0132] Example 2:

[0133] like Figure 2 As shown, this embodiment provides a non-motorized vehicle speeding recognition device based on image sequences. See [link to relevant documentation]. Figure 2 The device includes an acquisition unit 701, a mapping unit 702, a modeling unit 703, a processing unit 704, a removal unit 705, and an analysis unit 706.

[0134] The acquisition unit 701 is used to acquire image sequence data of non-motorized vehicles in low-light curved passages;

[0135] The mapping unit 702 is used to extract non-motorized vehicle boundary features and channel structure parameters based on the image sequence data, and to perform mapping processing based on the extracted features and channel structure parameters to obtain the set of embedded trajectory points of non-motorized vehicles in the curved channel;

[0136] Modeling unit 703 is used to construct a direction-sensitive tensor graph and perform multi-scale steering perception graph modeling processing on the embedded trajectory point set to obtain a multi-dimensional graph structure containing time dimension, spatial direction and steering curvature information.

[0137] Processing unit 704 is used to process the multidimensional graph structure through an angular graph attention mechanism and a convolutional trajectory transformation network to obtain a speed estimation vector sequence for non-motorized vehicles;

[0138] The removal unit 705 is used to perform acceleration filtering and compensation processing on the velocity estimation vector sequence to obtain a velocity sequence with the turning inertia error removed.

[0139] The analysis unit 706 is used to analyze the speed sequence for non-motorized vehicles to determine whether the non-motorized vehicles have engaged in speeding behavior on curves.

[0140] It should be noted that the specific manner in which each module performs its operation in the apparatus described in the above embodiments has been described in detail in the embodiments of the method, and will not be elaborated here.

[0141] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0142] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying non-motorized vehicle speeding based on image sequences, characterized in that, include: Acquire image sequence data of non-motorized vehicles in low-light curved passages; Based on the image sequence data, non-motorized vehicle boundary features and channel structure parameters are extracted, and based on the extracted features and channel structure parameters, mapping processing is performed to obtain the set of embedded trajectory points of non-motorized vehicles in the curved channel; The embedded trajectory point set is processed by constructing a direction-sensitive tensor graph and modeling a multi-scale steering perception graph to obtain a multi-dimensional graph structure containing information on time dimension, spatial direction and steering curvature. The multidimensional graph structure is processed by an angular graph attention mechanism and a convolutional trajectory transformation network to obtain a speed estimation vector sequence for non-motorized vehicles; The velocity estimation vector sequence is subjected to acceleration filtering and compensation processing to obtain a velocity sequence with the turning inertia error removed; The speed sequence is analyzed to determine whether the non-motorized vehicle is speeding on curves. The process includes extracting non-motorized vehicle boundary features and channel structure parameters based on the image sequence data, and performing mapping processing based on the extracted features and channel structure parameters, including: The image sequence data is processed by boundary tensor response extraction. By constructing a multi-frequency guided edge tensor response model of the non-motorized vehicle candidate region in each image frame of the image sequence data, a preliminary non-motorized vehicle boundary response map is obtained. The preliminary non-motorized vehicle boundary response map is subjected to local boundary curvature estimation. By performing multi-scale orientation fitting on the non-motorized vehicle boundary points and introducing a second-order guidance term in the strong turning region to suppress image noise interference, a set of boundary curvature points is obtained. The boundary curvature point set is subjected to channel structure-guided perspective geometry modeling. By utilizing the channel structure parameters of the image sequence data, the trajectory in the image frame is mapped from the pixel domain to the curvature space in a multi-stage manner to obtain the predicted motion path sequence of the non-motorized vehicle. The motion path sequence is processed by manifold embedding to obtain the set of embedded trajectory points of non-motorized vehicles in the curved channel; The process of constructing a direction-sensitive tensor graph and modeling a multi-scale turning-aware graph from the embedded trajectory point set includes: The embedded trajectory point set is subjected to local direction encoding processing, wherein the motion direction between adjacent trajectory points is represented as a directed edge in a tensor to obtain an initial direction tensor graph. The initial orientation tensor is subjected to spatial orientation consistency optimization processing, wherein the spatial neighborhood average field is introduced into the initial orientation tensor and it is smoothed and irregularly turned to change the region to obtain the optimized tensor structure. The optimized tensor graph structure is subjected to multi-scale steering response modeling. By performing tensor domain hierarchical modeling on the steering gradient information at different scales, the feature response corresponding to the preset steering mode is extracted to obtain the steering perception graph structure. The steering perception graph structure is subjected to multidimensional graph structure fusion processing. By fusing the trajectory timestamp, position coordinates, direction vector and local curvature encoding into high-order graph node features, a multidimensional graph structure containing time dimension, spatial direction and steering curvature information is obtained.

2. The non-motorized vehicle speeding identification method based on image sequences according to claim 1, characterized in that... The multidimensional graph structure is processed through an angular graph attention mechanism and a convolutional trajectory transformation network, including: The multidimensional graph structure is subjected to attention weight initialization processing guided by angular partial derivative. The angular offset of each node in the multidimensional graph structure within its directional neighborhood is used as the kernel weight value for attention calculation to obtain the graph attention initial weight matrix. The initial weight matrix of the graph attention is subjected to node neighborhood feature aggregation processing. By fusing direction, time and curvature information in the variable neighborhood of the node, an aggregated feature tensor is formed, and a set of node state vectors is obtained. The set of node state vectors is subjected to trajectory convolution transformation. By performing sliding window convolution operation based on direction sequence on the sequential time axis of the trajectory, and introducing channel curvature as a convolution weight adjustment factor, a convolution feature map reflecting the dynamics of the trajectory time series is obtained. The convolutional feature map is subjected to velocity vector analysis. By performing first and second derivative analysis on the node sequence in the convolutional output domain, the velocity estimation vector sequence of non-motorized vehicles is obtained.

3. The non-motorized vehicle speeding identification method based on image sequences according to claim 1, characterized in that... The velocity estimation vector sequence is subjected to acceleration filtering and compensation processing, including: The velocity estimation vector sequence is subjected to local acceleration response extraction processing. By calculating the acceleration at each time node in the velocity estimation vector sequence and using the direction switching rate as a suppression factor, the original acceleration vector sequence of the non-motorized vehicle is obtained. The original acceleration vector sequence is subjected to inertial disturbance identification processing. By setting a steering inertial detection interval within the high curvature segment of the embedded trajectory, the non-real velocity fluctuation region dominated by the turning trajectory structure is identified, and a set of inertial disturbance time period markers is obtained. The set of inertial disturbance time period markers is reconstructed using a speed smoothing sequence. By weighting the speed values ​​within the marked time periods, a preliminary speed smoothing sequence for non-motorized vehicles is obtained. The initial velocity smoothing sequence is subjected to global trend calibration, wherein the overall velocity curve is refitted under the premise of ensuring energy conservation to obtain a velocity sequence with the turning inertia error removed.

4. A non-motorized vehicle speeding recognition device based on image sequences, characterized in that, include: The acquisition unit is used to acquire image sequence data of non-motorized vehicles in low-light curved passages; The mapping unit is used to extract non-motorized vehicle boundary features and channel structure parameters based on the image sequence data, and to perform mapping processing based on the extracted features and channel structure parameters to obtain the set of embedded trajectory points of non-motorized vehicles in the curved channel; The modeling unit is used to construct a direction-sensitive tensor graph and perform multi-scale steering perception graph modeling processing on the embedded trajectory point set to obtain a multi-dimensional graph structure containing time dimension, spatial direction and steering curvature information. The processing unit is used to process the multidimensional graph structure through an angular graph attention mechanism and a convolutional trajectory transformation network to obtain a speed estimation vector sequence for non-motorized vehicles; The removal unit is used to perform acceleration filtering and compensation processing on the velocity estimation vector sequence to obtain a velocity sequence with the turning inertia error removed. The analysis unit is used to analyze the speed sequence for non-motorized vehicles to determine whether the non-motorized vehicles have engaged in speeding behavior on curves. The mapping unit includes: The first mapping subunit is used to perform boundary tensor response extraction processing on the image sequence data. By constructing a multi-frequency guided edge tensor response model of the non-motor vehicle candidate region in each image frame of the image sequence data, a preliminary non-motor vehicle boundary response map is obtained. The second mapping subunit is used to perform local boundary curvature estimation processing on the preliminary non-motorized vehicle boundary response map. By performing multi-scale orientation fitting on the non-motorized vehicle boundary points and introducing a second-order guidance term in the strong turning region to suppress image noise interference, a set of boundary curvature points is obtained. The third mapping subunit is used to perform channel structure-guided perspective geometry modeling on the boundary curvature point set. By using the channel structure parameters of the image sequence data to perform multi-stage mapping from the pixel domain to the curvature space on the trajectory in the image frame, the estimated motion path sequence of the non-motorized vehicle is obtained. The fourth mapping subunit is used to perform manifold embedding processing on the motion path sequence to obtain the set of embedded trajectory points of non-motorized vehicles in the curve channel; The modeling unit includes: The first modeling subunit is used to perform local direction encoding processing on the embedded trajectory point set, wherein the motion direction between adjacent trajectory points is represented as a directed edge in a tensor to obtain an initial direction tensor graph. The second modeling subunit is used to perform spatial orientation consistency optimization processing on the initial orientation tensor graph. Specifically, the orientation average field in the spatial neighborhood is introduced into the initial orientation tensor graph and it is smoothed and irregularly turned to change the region to obtain the optimized tensor graph structure. The third modeling subunit is used to perform multi-scale steering response modeling on the optimized tensor graph structure. By performing tensor domain hierarchical modeling on the steering gradient information at different scales, the feature response corresponding to the preset steering mode is extracted to obtain the steering perception graph structure. The fourth modeling subunit is used to perform multi-dimensional graph structure fusion processing on the steering perception graph structure. By fusing the trajectory timestamp, position coordinates, direction vector and local curvature encoding into high-order graph node features, a multi-dimensional graph structure containing time dimension, spatial direction and steering curvature information is obtained.

5. The non-motorized vehicle speeding recognition device based on image sequence according to claim 4, characterized in that, The processing unit includes The first processing subunit is used to perform attention weight initialization processing guided by angular partial derivative on the multidimensional graph structure. By using the angular offset of each node in the multidimensional graph structure in its directional neighborhood as the kernel weight value for attention calculation, the graph attention initial weight matrix is ​​obtained. The second processing subunit is used to perform node neighborhood feature aggregation processing on the graph attention initial weight matrix. By fusing direction, time and curvature information in the variable neighborhood of the node, an aggregated feature tensor is formed to obtain a set of node state vectors. The third processing subunit is used to perform trajectory convolution transformation on the set of node state vectors. By performing sliding window convolution operation based on direction sequence on the sequential time axis of the trajectory, and introducing channel curvature as a convolution weight adjustment factor, a convolution feature map reflecting the dynamics of the trajectory time series is obtained. The fourth processing subunit is used to perform velocity vector analysis on the convolutional feature map. By performing first-order and second-order derivative analysis on the node sequence in the convolutional output domain, the velocity estimation vector sequence of the non-motorized vehicle is obtained.

6. The non-motorized vehicle speeding recognition device based on image sequence according to claim 4, characterized in that, The removal unit includes: The first removal subunit is used to perform local acceleration response extraction processing on the velocity estimation vector sequence. By calculating the acceleration at each time node in the velocity estimation vector sequence and using the direction switching rate as a suppression factor, the original acceleration vector sequence of the non-motorized vehicle is obtained. The second removal subunit is used to perform inertial disturbance identification processing on the original acceleration vector sequence. By setting a steering inertial detection interval in the high curvature segment of the embedded trajectory, the non-real speed fluctuation area dominated by the turning trajectory structure is identified, and a set of inertial disturbance time period markers is obtained. The third removal subunit is used to perform speed smoothing sequence reconstruction processing on the inertial disturbance time period mark set. By weighting the speed values ​​within the mark period, a preliminary speed smoothing sequence of non-motorized vehicles is obtained. The fourth removal subunit is used to perform global trend calibration processing on the preliminary velocity smoothing sequence, wherein the overall velocity curve is refitted under the premise of ensuring energy conservation to obtain a velocity sequence with the turning inertia error removed.

Images