Vehicle trajectory deduction method for large-area monitoring blind spots

Abstract

The present invention provides a method for deducing vehicle driving trajectories for large-area monitoring blind spots, which belongs to the field of intelligent transportation technology. The present invention first locates the starting and ending sections of the target vehicle, and then establishes the structure of the Bayesian network with reference to the topological relationship of the traffic network. Combined with real-time traffic data such as vehicle speed, traffic volume, road width, and quantified influencing factors, the probability values ​​of each parent node and child node are calculated to obtain the Bayesian network parameters. Subsequently, by selecting the sections corresponding to the nodes with higher probability values, the effective connection of the sections in the road network structure is achieved, and finally the driving trajectory deduction of the target vehicle from the starting point to the end point is completed. The present invention highly restores the actual driving trajectory of vehicles in large-area monitoring blind spots and improves the degree of coordination of various types of traffic data.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation technology, and in particular relates to a vehicle driving trajectory deduction method for large-area monitoring blind spots. Background Art

[0002] With the advancement of smart transportation, a variety of advanced traffic perception technologies have been introduced and deployed in ever-increasing density, providing technical support for the effective acquisition of dynamic targets. However, with the acceleration of urbanization, traffic networks are becoming increasingly complex. Despite the many achievements of smart transportation, comprehensive coverage and accurate perception of urban road networks remain elusive. Furthermore, the complex traffic environment and diverse driver intentions have led to a significant increase in the number of vehicle paths within the traffic network, making it particularly difficult to accurately reproduce the trajectory of target vehicles within such a large and complex network.

[0003] The existing technology involves vehicle trajectory deduction technology based on vehicle re-identification algorithms. This technical solution first collects video data captured by traffic intersection surveillance cameras and obtains the corresponding traffic network data. It then detects vehicles and extracts vehicle information using a deep learning-based target detection model. It then uses a re-identification algorithm to detect the same vehicle appearing in different surveillance videos. The vehicle's position in the surveillance image is then mapped to the traffic network. The shortest path algorithm is used to obtain the optimal path between two adjacent positions in a time series. These paths are connected in sequence to form a complete driving trajectory, and the tracking trajectory is finally visualized. However, its shortcomings are as follows:

[0004] Vehicle trajectory deduction technology based on vehicle re-identification algorithms is suitable for areas where surveillance cameras are densely deployed. When a vehicle travels into a traffic area with a low density of surveillance cameras and a complex road network structure, it will be extremely difficult to use the above technical solution to deduce the vehicle's true and continuous driving trajectory.

[0005] Existing technology involves inferring dynamic target motion trajectories based on Bayesian networks. A Bayesian network model consists of nodes, directed edges, and network parameters. In path inference technology, directed edges represent the connections between road segments (from the starting segment node to the connected segment nodes), while network parameters represent the conditional probabilities of selecting different segments. The inference results describe the target's possible motion direction at the next moment. However, this approach has the following drawbacks:

[0006] The Bayesian network solution applied to vehicle trajectory deduction adopts a single influencing factor quantification method in the parameter solution process and ignores the influence degree of each factor, making it unable to reasonably reflect the driver's decision intention, resulting in low accuracy of this technical solution. Summary of the Invention

[0007] In response to the above-mentioned deficiencies in the prior art, the present invention provides a vehicle driving trajectory deduction method for large-area monitoring blind spots, which solves the following problems: it optimizes the deficiency of the target vehicle re-identification method in the prior art that is unable to cope with large-area monitoring blind spots; it improves the existing Bayesian network deduction method to deduce vehicle driving trajectories, and realizes the continuous deduction and prediction of the trajectory of target vehicles in large-area monitoring blind spots.

[0008] In order to achieve the above objectives, the present invention adopts a technical solution: a vehicle trajectory deduction method for large-area monitoring blind spots, comprising the following steps:

[0009] S1. Obtain a traffic information dataset of a large monitoring blind spot;

[0010] S2. Based on the traffic information dataset, the structure of the Bayesian network model is constructed;

[0011] S3. Determine the quantitative impact factor of the Bayesian network model;

[0012] S4. Using the weight coefficient estimation process, estimate the influence of the quantitative factors by time period to obtain the weight coefficient;

[0013] S5. Calculate the weighted Bayesian network model parameters based on the influencing quantification factors and weight coefficients to obtain the vehicle driving trajectory deduction results.

[0014] The beneficial effects of the present invention are as follows: the present invention first locates the starting and ending sections of the target vehicle, and then establishes the structure of the Bayesian network with reference to the topological relationship of the traffic network. Combined with real-time traffic data such as vehicle speed, traffic volume, road width, and quantified influencing factors, the probability values ​​of each parent node and child node are calculated, and the Bayesian network parameters are obtained. Subsequently, by selecting the sections corresponding to the nodes with higher probability values, the effective connection of the sections in the road network structure is achieved, and finally the driving trajectory of the target vehicle from the starting point to the end point is deduced. The present invention highly restores the actual driving trajectory of the vehicle and improves the degree of coordination of various types of traffic data.

[0015] Furthermore, the S1 is specifically:

[0016] Based on the large-scale monitoring blind spots, the traffic network is divided into multiple independent sections and numbered;

[0017] Extract the length and width of each traffic section;

[0018] The traffic flow and average speed of each road section at different time periods are recorded using surveillance videos to construct a traffic information dataset.

[0019] The beneficial effect of the above further scheme is that the above scheme constructs a traffic information dataset, including independent road sections divided based on traffic intersections and road end points, and records the length, width, traffic flow, average speed, etc. of each road section, providing data support for constructing the Bayesian network model structure and quantifying influencing factors.

[0020] Furthermore, the S2 is specifically:

[0021] Based on the road network topology structure in the traffic information dataset, the starting and ending road segments are used as the starting and target nodes of the Bayesian network respectively, and the nodes other than the starting and target nodes are used as intermediate nodes.

[0022] The Bayesian network model structure is constructed based on the connection relationship between the intermediate nodes assigned by the vehicle driving direction on each road section.

[0023] The above-mentioned further solution has the beneficial effect of directly mapping the road network topology into a Bayesian network structure. This allows the Bayesian network structure to be mapped to the actual road network, intuitively reflecting the trajectories of vehicles on the road network and facilitating the understanding and analysis of traffic flow changes. Furthermore, this solution effectively avoids the drawbacks of traditional structured learning algorithms, such as high computational complexity and lack of prior knowledge when constructing Bayesian network structures.

[0024] Furthermore, the impact quantification factors include:

[0025] Road grade:

[0026]

[0027] in, and Respectively represent i The width of the road segment before and after normalization, and Respectively represent the maximum and minimum widths of all road sections within the measurement area;

[0028] Congestion level:

[0029]

[0030]

[0031] in, represents the normalized calculation of the road congestion index, 、 Respectively The endpoint of the traffic congestion determination interval, Indicates the average driving speed of the road section. Indicates the number of measured vehicles, Indicates the length of the road section, Indicates vehicle Time to pass through the section;

[0032] Motion Vector:

[0033]

[0034] in, and Represents the motion vector of a road segment before and after normalization;

[0035] Vehicle density:

[0036]

[0037] in, and Respectively represent the road intersection The vehicle density values ​​before and after normalization on the road segment, and They respectively represent the maximum and minimum vehicle density in all connected sections of the same road intersection.

[0038] The above-mentioned further solution has the beneficial effect of using a mathematical model to estimate the factors that influence driver decision-making. A higher value for the influence factor indicates a higher likelihood that the driver will choose that road section. Compared to traditional hierarchical, unified estimation schemes, this solution uses corresponding mathematical models for different influence factors, providing a more accurate measure of driver decision-making.

[0039] Furthermore, the S4 includes the following steps:

[0040] S401, dividing the vehicle operation time into morning peak, afternoon peak, evening peak and non-peak periods;

[0041] S402, initializing the weight coefficients of the impact quantization factors of each time period;

[0042] S403: Calculating weighted Bayesian network parameters based on the impact quantification factors and the weight coefficients corresponding to the vehicle operation time periods;

[0043] S404, performing Bayesian network deduction based on the constructed Bayesian network model structure and weighted Bayesian network parameters;

[0044] S405: Determine whether the deduced path is consistent with the actual path. If so, obtain the optimal weight coefficient for the time period. Otherwise, adjust the weight coefficient affecting the quantization factor for the time period and return to S403.

[0045] The beneficial effects of this further solution are: it quantitatively analyzes the impact of different factors on driving decisions, improving the coordination of various traffic data. Furthermore, the optimized Bayesian network model improves model accuracy and enhances the ability to predict driver behavior decisions.

[0046] Furthermore, the S5 is specifically as follows:

[0047] Substitute the influencing quantization factor and weight coefficient into the expression of weighted Bayesian network parameters to solve the network parameters and obtain the selection probability of the road section corresponding to each node;

[0048] According to the structure of the Bayesian network model, the road sections with the highest probability of being selected are connected in sequence to obtain the vehicle driving trajectory deduction result.

[0049] The beneficial effect of the above further scheme is that the vehicle driving trajectory deduction results obtained by the above scheme comprehensively consider multiple influencing quantitative factors such as traffic conditions, road properties, traffic flow, etc. Compared with traditional deduction methods, they can reflect the vehicle's driving path selection in actual traffic scenarios, and more accurately restore the vehicle's driving trajectory under different traffic conditions, providing a powerful tool for traffic management and analysis.

[0050] Furthermore, the expression of the selection probability is as follows:

[0051]

[0052] in, Indicates selection i The probability value of the road segment, 、 、 、 and All said 、 、 、 and The weight coefficient affecting the quantitative factor, Indicates the i Factors other than the quantitative factors affecting the road section.

[0053] The beneficial effect of the above further scheme is: the above scheme is the core of the weighted Bayesian network parameter calculation, and the weight coefficient is introduced on the basis of the traditional Bayesian network model, which effectively improves the model accuracy and the synergistic relationship of various influencing quantitative factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Flow chart of the method of the present invention.

[0055] Figure 2This is a schematic diagram of the experimental measurement area and road section numbering.

[0056] Figure 3 Schematic diagram of the road network structure and its corresponding Bayesian network.

[0057] Figure 4 Schematic diagram of the deduction results using this invention. DETAILED DESCRIPTION

[0058] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0059] Example

[0060] like Figure 1 As shown, the present invention provides a vehicle driving trajectory deduction method for large-area monitoring blind spots, and its implementation method is as follows:

[0061] S1. Obtain a traffic information dataset for a large-area blind spot, specifically:

[0062] Based on the large-scale monitoring blind spots, the traffic network is divided into multiple independent sections and numbered;

[0063] Extract the length and width of each traffic section;

[0064] The traffic flow and average speed of each road section at different time periods are recorded using surveillance videos to construct a traffic information dataset.

[0065] S2. Based on the traffic information dataset, the structure of the Bayesian network model is constructed, which is specifically as follows:

[0066] Based on the road network topology structure in the traffic information dataset, the starting and ending road segments are used as the starting and target nodes of the Bayesian network respectively, and the nodes other than the starting and target nodes are used as intermediate nodes.

[0067] The Bayesian network model structure is constructed based on the connection relationship between the intermediate nodes assigned by the vehicle driving direction on each road section.

[0068] S3. Determine the quantitative impact factor of the Bayesian network model;

[0069] In this embodiment, the impact quantification factor constitutes the key parameter in the Bayesian network model. Based on the behavior pattern of ordinary drivers, the present invention selects factors such as road grade, congestion level, motion vector and vehicle density as key influencing factors for drivers to choose the optimal road section in urban traffic environment.

[0070] (1) Road grade: Select road section width data As the basis for judging the road grade, the road grade of the present invention is obtained by normalizing the road section width using formula (1): .when The larger it is, the wider the road is, and the greater the probability that the driver will prefer this road section.

[0071] (1)

[0072] in, and Respectively represent i The width of the road segment before and after normalization, and They respectively represent the maximum and minimum widths of all road sections within the measurement area.

[0073] (2) Congestion level: Congestion index It is an important indicator that comprehensively reflects road congestion. Its value range is 0-10. The higher the value, the more serious the traffic congestion. Therefore, the congestion index selected in this invention is As the basis for judging the degree of congestion. Use formula (2) to calculate the average driving speed of the road section, and refer to Table 1 to achieve the urban traffic congestion degree and congestion index Table 1 shows the corresponding relationship between the average driving speed of urban road sections and the traffic congestion index.

[0074] (2)

[0075] in, Indicates the average driving speed of the road section. Indicates the number of measured vehicles, Indicates the length of the road section, Indicates vehicle The time it takes to pass through the section.

[0076] Table 1

[0077]

[0078] The present invention uses formula (3) to realize the normalized calculation of the road congestion index based on the average driving speed of the road section and the traffic congestion determination interval in which it is located. The specific calculation is shown in formula 3:

[0079] (3)

[0080] in, represents the normalized calculation of the road congestion index, 、 Respectively The endpoint of the traffic congestion determination interval, Indicates the average driving speed of the road section. Indicates the number of measured vehicles, Indicates the length of the road section, Indicates vehicle Time to pass through the section; The larger the value, the lower the congestion level, and the greater the probability that the driver will prefer this road section. Set to 1.

[0081] (3) Motion vector: At a road intersection, when drivers choose the next driving section, they usually prefer to choose a route close to the destination. In order to evaluate the correlation between the vehicle's driving direction and the destination direction, the present invention uses the concept of "motion vector" to analyze whether the vehicle is moving towards the destination. By quantifying the angle between the vehicle's driving direction and the target direction (the direction of the line connecting the starting point and the destination), the vehicle is detected. (in ), can effectively judge the vehicle's driving intention. When it is smaller, it indicates that the target vehicle is closer to the destination; otherwise, it indicates that the vehicle may deviate from the destination or take a detour.

[0082] According to the inverse relationship between the motion direction angle and the driving intention, the present invention uses the cosine function to normalize the angle to obtain the normalized motion vector of a certain road section. , see formula (4) for the specific method. When it is close to 0, it indicates that the road section selected by the target vehicle deviates significantly from the destination direction; When it is close to 1, it means that the target vehicle is moving closer to the destination.

[0083] (4)

[0084] in, and Represent the motion vector of a road section before and after normalization respectively.

[0085] (4) Vehicle density: Vehicle density refers to the density of vehicles on a road. Generally, drivers tend to choose road sections with lower vehicle density, which not only improves driving comfort, provides more room to deal with danger in emergencies, but also effectively avoids traffic congestion during peak hours. See formula (2). Vehicle density can be calculated by the ratio of the traffic flow of a given road section to the average driving speed of the section.

[0086] At road intersections, drivers often make quick decisions about their driving paths by comparing vehicle density values ​​across different road sections. To reflect this decision-making process, the present invention first calculates the vehicle density values ​​for each road section connected to the intersection. Subsequently, the vehicle density values ​​for different road sections at the same intersection are normalized, as shown in Formula (5):

[0087] (5)

[0088] in, and Respectively represent the road intersection The vehicle density values ​​before and after normalization on the road segment, and Respectively represent the maximum and minimum vehicle density in all connected sections of the same road intersection. The closer it is to 1, the more drivers are willing to choose this road section.

[0089] S4. Use the weight coefficient estimation process to estimate the influence of the quantitative factors by time period and obtain the weight coefficient. The implementation method is as follows:

[0090] S401, dividing the vehicle operation time into morning peak, afternoon peak, evening peak and non-peak periods;

[0091] S402, initializing the weight coefficients of the impact quantization factors of each time period;

[0092] S403: Calculating weighted Bayesian network parameters based on the impact quantification factors and the weight coefficients corresponding to the vehicle operation time periods;

[0093] S404, performing Bayesian network deduction based on the constructed Bayesian network model structure and weighted Bayesian network parameters;

[0094] S405: Determine whether the deduced path is consistent with the actual path. If so, obtain the optimal weight coefficient for the time period. Otherwise, adjust the weight coefficient affecting the quantization factor for the time period and return to S403.

[0095] S5. Calculate the weighted Bayesian network model parameters based on the influencing quantification factors and weight coefficients to obtain the vehicle trajectory deduction results, which are specifically:

[0096] Substitute the influencing quantization factor and weight coefficient into the expression of weighted Bayesian network parameters to solve the network parameters and obtain the selection probability of the road section corresponding to each node;

[0097] According to the structure of the Bayesian network model, the road sections with the highest probability of being selected are connected in sequence to obtain the vehicle driving trajectory deduction result.

[0098] In this embodiment, the present invention first locates the starting and ending road sections of the target vehicle based on a vehicle re-identification algorithm, and then establishes the Bayesian network structure by referring to the traffic network topology. Combining real-time traffic data such as vehicle speed, traffic volume, and road width, and based on quantitative influencing factors, formula (6) is used to calculate the probability values ​​of each parent node and child node, thereby obtaining the Bayesian network parameters. Subsequently, by selecting the road sections corresponding to the nodes with higher probability values, the effective connection of the road sections in the road network structure is achieved, and the driving trajectory of the target vehicle from the starting point to the end point is finally deduced.

[0099] (6)

[0100] in, i Indicates the i road section, , Indicates the n road section, Indicates the i The number of factors affecting the road section in addition to the quantitative factors, 、 、 、 and All said 、 、 、 and The estimation process of the weight coefficients affecting the quantitative factors refers to the idea of ​​sensitivity analysis and is obtained through gradual iterative optimization during the model deduction process.

[0101] The present invention is subjected to simulation experiments as follows.

[0102] (1) Data source

[0103] In this embodiment, the campus road network is selected as the experimental sample area, and a campus road network vector map is constructed based on satellite remote sensing images. The traffic network in the experimental area is divided into multiple independent sections and numbered using traffic intersections and road ends as dividing points. Figure 2 As shown, the black arrow indicates the direction in which the road is passable.

[0104] In this embodiment, high-precision remote sensing images are used to extract the length and width of each traffic section, and surveillance video data is used to record the traffic flow and average speed of each section at different time periods, thereby constructing the traffic information dataset required for the experiment in this paper. The data record is shown in Table 1 as an example.

[0105] Table 1

[0106]

[0107] (2) Bayesian network model structure construction

[0108] The experiment is based on the vehicle re-identification technology extracted from the campus road network. For example, a vehicle starts driving westward from section 01 one night and finally leaves southward from section 12. The experiment uses the starting section and the ending section as the starting nodes of the Bayesian network. With the target node , and the remaining independent road sections in the experimental area are used as intermediate nodes Based on the driving direction of vehicles on each road section, the connection relationship between each intermediate node is given, and the Bayesian network structure constructed is as follows Figure 3 As shown, Figure 3 (a) is a schematic diagram of the road network structure. Figure 3 (b) is a schematic diagram of the Bayesian network structure.

[0109] (3) Bayesian network model structure construction

[0110] The experiment selected four influencing factors—road grade, congestion level, motion vector, and vehicle density—to reflect driver decision intentions. Based on the road width, road length, traffic flow, and average speed data from the traffic information dataset in Table 1, the aforementioned method was used to quantify these influencing factors, resulting in the quantified influencing factors for each node in Table 2.

[0111] Table 2

[0112]

[0113] (4) Network parameter calculation

[0114] In this embodiment, in order to measure the influence of different influencing factors on the driver's decision, the weight coefficient estimation process is used to obtain the time-divided influencing factor weight coefficients shown in Table 3, and then the time-divided weighted Bayesian network parameter calculation formula is obtained, as shown in Formula (7).

[0115] Table 3

[0116]

[0117] (7)

[0118] (5) Deduction results

[0119] Substitute the quantified impact factor into Equation 7 (late) to solve the network parameters and obtain Figure 3 The probability distribution of the road segments corresponding to each node in (b) is shown in Table 4, which is a table of vehicle trajectory deduction results. According to the Bayesian network structure, the road segments with the highest probability of selection are connected in sequence, and the deduction path result of the present invention is obtained as follows: .

[0120] Table 4

[0121]

[0122] (6) Result verification

[0123] This experiment collected surveillance video data captured by fixed cameras on campus. There were no blind spots in the surveillance video within the measurement area. Based on vehicle re-identification technology, the same vehicle appearing in different videos was extracted, and the video clips of the target vehicle appearing in the surveillance were recorded in sequence, thus forming a total of 100 complete driving trajectories for deduction experiments and result verification.

[0124] The starting and ending sections of the actual path are set as known conditions, and the sections in the gray area are assumed to be monitoring blind spots. The present invention is used to conduct experiments, and the deduction experimental results are compared with the actual trajectory in the video clip to verify the accuracy of the model results. Taking the above-mentioned target vehicle as an example, the deduction results obtained based on the present invention are as follows: Figure 4 As shown. The red arrow is the driving trajectory with the highest probability obtained by the scheme in this paper; the percentage after the road section number is the probability distribution of the road section being selected; the video clip is the verification point of whether the vehicle passes through the road section. Specifically, the target vehicle is firstly Surveillance video recognition, appearing in sequence 、 、 Verification point No., finally at the end of the road section The surveillance video recognition and deduction results are completely consistent with the actual path in the video clip.

Claims

1. A vehicle trajectory deduction method for large-area monitoring blind spots, characterized by: The following steps are involved: S1. Obtain a traffic information dataset of a large monitoring blind spot; S2. Based on the traffic information dataset, the structure of the Bayesian network model is constructed; S3. Determine the impact quantification factor of the Bayesian network model, where the impact quantification factor includes road grade , congestion level , motion vector and vehicle density ; S4. Using the weight coefficient estimation process, estimate the influence of the quantitative factors by time period to obtain the weight coefficient; S5. Calculate the weighted Bayesian network model parameters based on the influencing quantification factors and weight coefficients to obtain the vehicle trajectory deduction results, which are specifically: Substitute the influencing quantization factor and weight coefficient into the expression of weighted Bayesian network parameters to solve the network parameters and obtain the selection probability of the road section corresponding to each node; According to the structure of the Bayesian network model, the road sections with the highest probability of being selected are connected in sequence to obtain the vehicle trajectory deduction result; The expression of the selection probability is as follows: in, Indicates selection i The probability value of the road segment, 、 、 、 and Respectively 、 、 、 and The weight coefficient affecting the quantitative factor, Indicates the i Factors other than the quantitative factors affecting the road section.

2. The vehicle trajectory deduction method for large-area monitoring blind spots according to claim 1 is characterized in that: The S1 is specifically: Based on the large-scale monitoring blind spots, the traffic network is divided into multiple independent sections and numbered; Extract the length and width of each traffic section; The traffic flow and average speed of each road section at different time periods are recorded using surveillance videos to construct a traffic information dataset.

3. The vehicle driving trajectory deduction method for large-area monitoring blind spots according to claim 1 is characterized in that: The S2 is specifically: Based on the road network topology structure in the traffic information dataset, the starting and ending road segments are used as the starting and target nodes of the Bayesian network respectively, and the nodes other than the starting and target nodes are used as intermediate nodes. The Bayesian network model structure is constructed based on the connection relationship between the intermediate nodes assigned by the vehicle driving direction on each road section.

4. The vehicle trajectory deduction method for large-area monitoring blind spots according to claim 1 is characterized in that the road grade: in, and Respectively represent i The width of the road segment before and after normalization, and Respectively represent the maximum and minimum widths of all road sections within the measurement area; Congestion level: in, represents the normalized calculation of the road congestion index, 、 Respectively The endpoint of the traffic congestion determination interval, Indicates the average driving speed of the road section. Indicates the number of measured vehicles, Indicates the length of the road section, Indicates vehicle Time to pass through the section; Motion Vector: in, and Represents the motion vector of a road segment before and after normalization; Vehicle density: in, and Respectively represent the road intersection The vehicle density values ​​before and after normalization on the road segment, and They respectively represent the maximum and minimum vehicle density in all connected sections of the same road intersection.

5. The vehicle trajectory deduction method for large-area monitoring blind spots according to claim 4 is characterized in that: The S4 comprises the following steps: S401, dividing the vehicle operation time into morning peak, afternoon peak, evening peak and non-peak periods; S402, initializing the weight coefficients of the impact quantization factors of each time period; S403: Calculating weighted Bayesian network parameters based on the impact quantification factors and the weight coefficients corresponding to the vehicle operation time periods; S404, performing Bayesian network deduction based on the constructed Bayesian network model structure and weighted Bayesian network parameters; S405: Determine whether the deduced path is consistent with the actual path. If so, obtain the optimal weight coefficient for the period. Otherwise, adjust the weight coefficient of the influencing quantization factor for the period and return to S403.

Images