A machine learning based road surface condition detection system

Abstract

This invention discloses a road condition detection system based on machine learning, comprising an infrared data acquisition unit, an execution unit, a target discrimination unit, a risk inference unit, and a level determination unit. The infrared data acquisition unit is constructed based on an infrared beam array and can continuously monitor the ambient light intensity and the signal-to-noise ratio of the received signal, dynamically calculating and outputting new transmission power and wavelength switching commands based on preset thresholds. The execution unit can individually and systematically control each infrared transmitter group based on a preset scanning encoding matrix and wavelength switching commands, determining the identifier of each infrared beam obscured by debris within the scanning cycle. This invention utilizes an infrared beam array unit deployed close to the road surface as the acquisition unit, integrating adaptive wavelength switching and transmission power regulation, facilitating the maintenance of stable signal quality and detection sensitivity under complex lighting and weather conditions.

Description

Technical Field

[0001] This invention relates to the field of road condition detection technology, and more specifically to a road condition detection system based on machine learning. Background Technology

[0002] Road debris, including dropped goods, tire fragments, and construction waste, poses a significant threat to road safety. These obstacles not only directly encroach on driving space but also force drivers to swerve at high speeds, easily leading to rear-end collisions, rollovers, and other traffic accidents, seriously threatening road safety. Therefore, using a road surface condition detection system is of great practical significance for preventing such accidents and protecting life and property.

[0003] For example, the machine learning-based road surface condition detection system and method disclosed in Chinese patent publication number CN 117079149 B discloses a method that changes the existing periodic detection method by constructing changes in road surface condition based on the influence of changes in road type, road load, meteorological factors, etc., and adjusts the detection cycle of road surface condition detection in real time in order to detect problems in road surface condition in a timely manner, so as to ensure driving safety and road surface to achieve good driving conditions.

[0004] However, most road condition detection systems currently rely on a single sensing method, are highly dependent on visible light cameras, and are easily affected by lighting conditions and weather conditions. At the same time, most of these systems are based on static image recognition technology, which has a weak ability to process and warn of dynamic spills. Summary of the Invention

[0005] Therefore, the present invention provides a road surface condition detection system based on machine learning to solve the problems in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A machine learning-based road condition detection system includes an infrared data acquisition unit, an execution unit, a target discrimination unit, a risk inference unit, and a level determination unit.

[0008] The infrared data acquisition unit is built on an infrared beam array and can continuously monitor the ambient light intensity and the signal-to-noise ratio of the receiver signal. It can also dynamically calculate and output new transmission power and wavelength switching commands based on preset thresholds.

[0009] The execution unit can perform individual and orderly control on each infrared emitter group based on a preset scanning encoding matrix and wavelength switching instructions, and determine the identifier of each infrared beam blocked by the projectile within the scanning cycle.

[0010] The target discrimination unit can calculate the coordinates (x, y) of the interruption point based on the infrared beam identifier, and associate and cluster them based on the temporal and spatial proximity of the interruption point to form suspected targets, and calculate the confidence score of the suspected targets. Identify the scattered material;

[0011] The risk simulation unit can simulate and determine the trajectory of the spilled material based on real-time traffic flow data, and calculate the dynamic risk value between the spilled material and vehicles. ;

[0012] Dynamic risk value The calculation formula is as follows:

[0013]

[0014] in, Let be the collision probability. The probability of secondary accidents. Dynamic weights;

[0015] The level determination unit can determine dynamic risk values. With preset multi-level risk thresholds By comparing data, the instantaneous threat level of the spilled material can be dynamically determined.

[0016] Furthermore, the specific details of the infrared data acquisition unit are as follows:

[0017] 1) Based on the infrared beam-beam array, the array can continuously sample the ambient light intensity E at the millisecond level;

[0018] 2) Compare the current light intensity E with the preset threshold. Compare;

[0019] 3) Dynamically calculate and adjust the transmit power based on the illumination intensity E. This enables the drive circuit to adjust according to the transmission power. The value adjusts the emission intensity of the infrared beam array; the calculation formula is as follows:

[0020]

[0021] in, Based on the transmission power, and All are preset weighting coefficients. To calibrate ambient light reference values, The signal-to-noise ratio of the currently received signal. The target signal-to-noise ratio is preset.

[0022] Furthermore, the target discrimination unit includes a calculation subunit, a data processing subunit, and a discrimination subunit;

[0023] The computing subunit is able to receive the infrared beam identifier and calculate the two-dimensional coordinates (x,y) of the interruption point in real time based on the beam space equation and timestamp.

[0024] The data processing subunit can associate and cluster interruption points within a time period based on a machine learning spatiotemporal fusion network to generate a set of suspected targets.

[0025] The discrimination subunit can extract features from each of the suspected targets to form an infrared feature vector. Then based on infrared feature vectors Calculate confidence score .

[0026] Furthermore, the confidence score The calculation formula is as follows:

[0027]

[0028] in, It is the Sigmoid activation function. This is a feature embedding function used to embed infrared feature vectors. Mapped to 64-dimensional space, These are the weighting coefficients. This is the first bias term.

[0029] Furthermore, the discrimination subunit can also make real-time judgments based on the occlusion pattern of the infrared beam.

[0030] Furthermore, the specific content of the risk simulation unit is as follows:

[0031] 1) Import the real-time digital twin scene of the road segment and inject the real-time traffic flow data into the real-time digital twin scene;

[0032] 2) A virtual traffic flow is generated based on a probabilistic simulation algorithm and confidence scores, and the detection interval is determined by the initial positions of all vehicles and the coordinates of the spilled material;

[0033] 3) Aggregate all collision risks within the detection range and output dynamic risk values. .

[0034] Furthermore, the collision probability The calculation formula is as follows:

[0035]

[0036] in, For vehicles at any time Longitudinal distance from the spilled material and All are preset weighting coefficients. For action instruction function, This represents the probability of historical collisions.

[0037] Furthermore, the probability of the secondary accident The calculation formula is as follows:

[0038]

[0039] in, As the weight for accident type, Indicate whether there are other vehicles around the vehicle. This is the real-time traffic density coefficient.

[0040] Furthermore, if other vehicles are present around the vehicle, then If there are no other vehicles around the vehicle, then .

[0041] This invention has the following advantages: By deploying an infrared beam array unit close to the road surface as the acquisition unit and integrating wavelength adaptive switching and transmission power control, it fundamentally changes the shortcomings of traditional solutions that rely on passive ambient light. The system can dynamically select either 850nm or 940nm operating wavelength based on real-time ambient light intensity to avoid interference, thereby maintaining stable signal quality and detection sensitivity under complex lighting and weather conditions such as darkness, strong light, rain, and fog, achieving effective detection of road debris.

[0042] Meanwhile, once the spilled material is identified, digital twin technology can simulate future traffic conflicts in a virtual scenario in real time, transforming the static spatial information of the spilled material into a quantified dynamic risk value and providing graded early warnings, thereby improving the system's proactive early warning and accident prevention capabilities.

[0043] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0044] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0045] Figure 1This is a block diagram of a machine learning-based road condition detection system according to the present invention. Detailed Implementation

[0046] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Technical engineers in the field can make some non-essential improvements and adjustments to the present invention based on the above-described content. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Please see Figure 1 A machine learning-based road condition detection system includes an infrared data acquisition unit, an execution unit, a target discrimination unit, a risk inference unit, and a level determination unit.

[0048] The infrared data acquisition unit is built upon an infrared beam-and-beam array. It continuously monitors the ambient light intensity and the signal-to-noise ratio (SNR) of the received signal. Combined with preset thresholds, it dynamically calculates and outputs new transmit power and wavelength switching commands, ensuring that the receiver receives a stable signal with a suitable SNR under any external lighting conditions, providing a reliable data foundation for subsequent detection. It dynamically counteracts environmental interference, providing usable infrared signals for the system.

[0049] The specific contents of the infrared data acquisition unit are as follows:

[0050] 1) The infrared data acquisition unit is constructed based on infrared beam arrays deployed on both sides of the road and close to the road surface, which can continuously sample the ambient light intensity E at the millisecond level.

[0051] First, an infrared beam is emitted by an infrared transmitter deployed on one side of the road; when the beam passes through the road space and is captured by a receiver on the other side, the infrared data acquisition unit converts the received light signal into an electrical signal.

[0052] 2) Compare the current light intensity E with the preset threshold. Compare them. If E < This indicates that the ambient light intensity is low, and the interference from sunlight or urban lighting in the 850nm band is relatively small. Therefore, the control relay switches to the 850nm infrared beam array to act as the signal transmission source; if E ≥ When the ambient light intensity on the surface is high, there may be strong interference from sunlight or urban lighting in the 850nm band, while the interference in the 940nm band is relatively small. Therefore, the control relay switches to the 940nm infrared beam array to avoid strong interference from sunlight or urban lighting in a specific band.

[0053] 3) Dynamically calculate and adjust the transmit power based on the illumination intensity E. This enables the drive circuit to adjust according to the transmission power. The emission intensity of the infrared beam array is adjusted using pulse width modulation (PWM) or current control. The calculation formula is as follows:

[0054]

[0055] in, Based on the transmission power, and All are preset weighting coefficients. To calibrate ambient light reference values, The signal-to-noise ratio of the currently received signal. The target signal-to-noise ratio is preset.

[0056] Signal-to-noise ratio of the currently received signal The calculation formula is as follows:

[0057]

[0058] in, The average power of the infrared signal. This is environmental noise.

[0059] The execution unit can control each infrared emitter group individually and in an orderly manner based on a preset scanning coding matrix and wavelength switching instructions. It can identify the infrared beam identifier of each beam blocked by the scattering object within the scanning cycle, and conveniently and orderly activate sparse and non-adjacent infrared emitter groups. This allows the infrared beams to scan the road surface in turn according to a predetermined spatial path, effectively isolating each beam in space and time, and fundamentally avoiding the mutual interference caused by traditional multi-beam constant illumination.

[0060] In the infrared beam array, each infrared transmitter group is divided into several sub-units with an 8-meter interval along the road section. Each sub-unit is driven by the timing command of the scanning coding matrix to simultaneously and cyclically emit infrared transmitters one by one. When the projectile enters the detection area and blocks the beam, multiple beam signals with specific spatial orientations that are blocked at different times are recorded to obtain infrared beam identifiers, which facilitates the improvement of the response speed of projectile identification and the reduction of identification time.

[0061] The row index in the scanning coding matrix corresponds to different infrared emitter groups, and the column index corresponds to time points. This means that at any given time point, the physical locations of the activated and emitted beams within the entire sub-unit are sparse and non-adjacent, thus fundamentally avoiding the signal crosstalk and path confusion problems caused by multiple infrared beams forming a dense cross network in space.

[0062] The target discrimination unit can calculate the coordinates (x, y) of the interruption point based on the infrared beam identifier, and associate and cluster them based on the temporal and spatial proximity of the interruption point to form suspected targets, and calculate the confidence score of the suspected targets. Identify the object to be disposed of.

[0063] The target discrimination unit includes a calculation subunit, a data processing subunit, and a discrimination subunit.

[0064] The computational subunit can receive infrared beam identifiers and perform real-time calculations based on the beam space equation and timestamps. When a projectile obstructs the beam, the system records the unique identifier of the obstructed beam, the precise timestamp of the interruption, and then calculates the two-dimensional coordinates (x, y) of the interruption point using the beam space equation. The beam space equation is as follows:

[0065]

[0066] in, For the first The coordinates of the fixed emission point of the beam. For the first The detection radius of a single beam.

[0067] The data processing subunit can associate and cluster interruptions within a time period based on a machine learning spatiotemporal fusion network. The machine learning spatiotemporal fusion network can utilize the temporal and spatial proximity of interruptions to initially aggregate multiple interruptions belonging to the same physical object, generating a set of suspected targets.

[0068] The aforementioned spatiotemporal fusion network of machine learning refers to a class of algorithmic frameworks that can collaboratively process time series and spatially distributed data. It is mainly used to perform correlation analysis and clustering on asynchronously arriving data points with spatiotemporal labels (such as infrared beam interruption events) to identify potential spatial entities or events. In terms of specific implementation, it can be an attention-based algorithm such as ASTF-Net or graph neural network designed specifically for this task. The purpose is to reliably aggregate discrete observation points into a meaningful set of suspected targets based on the spatiotemporal proximity criterion.

[0069] The discrimination subunit can extract features from each suspected target, including the ID of the interrupted beam, the coordinates of the interruption point, and the interruption timestamp, to form an infrared feature vector. , Then based on infrared feature vectors Calculate confidence score ;when When the threshold is reached, the suspected target is confirmed as a real spill event, which facilitates the use of confidence scores. This reflects the probability that a suspected target is actually a piece of spilled material. Confidence score The calculation formula is as follows:

[0070]

[0071] in, It is the Sigmoid activation function. This is a feature embedding function used for infrared feature vectors. Mapping to a 64-dimensional space enhances the expressive power of features. These are the weighting coefficients. This is the first bias term. The embedding formula is as follows:

[0072]

[0073] in, This is the weight matrix. The matrix has 64 rows and 10 columns. , For activation function, This is the second bias term.

[0074] The discrimination subunit can also make real-time judgments based on the obstruction pattern of the infrared beams. For example, single beam obstruction triggers a level one alarm, marks it as a suspected target, and increases the monitoring frequency; if three adjacent beams are obstructed in a very short period of time, it is directly determined to be a real spill and triggers a level two linkage response; if multiple areas of beams are obstructed simultaneously, it is predicted to be a large-scale spill event, and the highest level of emergency control is initiated.

[0075] The risk simulation unit can simulate and determine the trajectory of spilled materials based on real-time traffic flow data, and calculate the dynamic risk value between the spilled materials and vehicles. This facilitates the transformation of static spatial information of spilled materials into dynamic assessments of collision risks, providing a direct basis for subsequent graded early warning systems.

[0076] The specific content of the risk simulation unit is as follows:

[0077] 1) Import the real-time digital twin scene of the road segment and inject real-time traffic flow data into the real-time digital twin scene. The real-time traffic flow data includes average vehicle speed, lane-level traffic density ρ, and weather conditions.

[0078] 2) Virtual traffic flow is generated based on probabilistic simulation algorithms and confidence scores, which facilitates subsequent simulation of vehicle avoidance behavior and deduction of collision risks;

[0079] First, based on real-time traffic density and road segment length Determine the number of virtual vehicles in the simulation. , This ensures that the density of virtual traffic flow is consistent with reality;

[0080] Then, random sampling is performed on the virtual vehicles at their initial speeds, using the real-time average speed as the mean and the standard deviation determined by historical statistical data, to simulate natural speed fluctuations. The detection range is determined by the initial positions of all vehicles and the coordinates of the spilled material.

[0081] The aforementioned probabilistic simulation algorithms refer to a class of computational methods that simulate the behavior of uncertain systems based on random sampling and statistical models. They are mainly used to generate virtual vehicles and their initial motion states that conform to real-world statistical laws based on real-time traffic flow parameters (density, average vehicle speed), thereby constructing a high-fidelity simulation input environment for subsequent collision risk probability deduction. In this field, other algorithms that can be used to achieve similar functions include traffic flow models based on deterministic rules (such as car-following models) and multi-agent simulation (ABM).

[0082] 3) Aggregate all collision risks within the detection range and output dynamic risk values. This supports tiered early warning and decision-making. Dynamic risk value. The calculation formula is as follows:

[0083]

[0084] in, Let be the collision probability. The probability of secondary accidents. Dynamic weights. Collision probability. The calculation formula is as follows:

[0085]

[0086] in, For vehicles at any time Longitudinal distance from the spilled material and All are preset weighting coefficients. This is an action indicator function, taking the value 0 or 1. This represents the probability of historical collisions.

[0087] Probability of secondary accidents The calculation formula is as follows:

[0088]

[0089] in, As the weight for accident type, This indicates whether there are other vehicles around the vehicle. If there are other vehicles around the vehicle, then... If there are no other vehicles around the vehicle, then . This is the real-time traffic density coefficient.

[0090] Dynamic weights The calculation formula is as follows:

[0091]

[0092] in, The vehicle's current speed. This is the maximum speed limit for the road. This is the critical distance threshold between the vehicle and the spilled material.

[0093] The level determination unit can determine the dynamic risk value With preset multi-level risk thresholds Comparison and dynamic determination of the instantaneous threat level of the spilled material: when < When the risk is low, only the event is recorded; when... ≤ < When the risk level is determined to be medium, a general warning is triggered; when ≤ < When a high-risk situation is identified, proactive early warning and traffic intervention are initiated; when ≥ When a situation is deemed a critical risk, the highest level of emergency response is immediately implemented, thereby mapping the quantified risk value into specific, tiered early warning and control instructions.

[0094] This invention utilizes an infrared beam array unit deployed close to the road surface as the data acquisition unit, integrating wavelength adaptive switching and transmission power control. This fundamentally overcomes the shortcomings of traditional solutions that rely on passive ambient light. The system can dynamically select either 850nm or 940nm operating wavelengths based on real-time ambient light intensity to avoid interference, thus maintaining stable signal quality and detection sensitivity under complex lighting and weather conditions such as darkness, strong light, rain, and fog, achieving effective detection of road debris.

[0095] Meanwhile, once the spilled material is identified, digital twin technology can simulate future traffic conflicts in a virtual scenario in real time, transforming the static spatial information of the spilled material into a quantified dynamic risk value and providing graded early warnings, thereby improving the system's proactive early warning and accident prevention capabilities.

[0096] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A road surface condition detection system based on machine learning, characterized in that, It includes an infrared data acquisition unit, an execution unit, a target discrimination unit, a risk simulation unit, and a level determination unit; The infrared data acquisition unit is built on an infrared beam array and can continuously monitor the ambient light intensity and the signal-to-noise ratio of the receiver signal. It can also dynamically calculate and output new transmission power and wavelength switching commands based on preset thresholds. The execution unit can perform individual and orderly control on each infrared emitter group based on a preset scanning encoding matrix and wavelength switching instructions, and determine the identifier of each infrared beam blocked by the projectile within the scanning cycle. The target discrimination unit can calculate the coordinates (x, y) of the interruption point based on the infrared beam identifier, and associate and cluster them based on the temporal and spatial proximity of the interruption point to form suspected targets, and calculate the confidence score of the suspected targets. Identify the scattered material; The risk simulation unit can simulate and determine the trajectory of the spilled material based on real-time traffic flow data, and calculate the dynamic risk value between the spilled material and vehicles. ; Dynamic risk value The calculation formula is as follows: in, Let be the collision probability. This represents the probability of secondary accidents. Dynamic weights; The level determination unit can determine the dynamic risk value. With preset multi-level risk thresholds By comparing data, the instantaneous threat level of the spilled material can be dynamically determined.

2. The road surface condition detection system based on machine learning according to claim 1, characterized in that, The specific details of the infrared data acquisition unit are as follows: 1) The ambient light intensity E is continuously sampled at the millisecond level using an infrared beam-beam array; 2) Compare the current light intensity E with the preset threshold. Compare; 3) Dynamically calculate and adjust the transmit power based on the illumination intensity E. This enables the drive circuit to adjust according to the transmission power. The value adjusts the emission intensity of the infrared beam array; the calculation formula is as follows: in, Based on the transmission power, and All are preset weighting coefficients. To calibrate ambient light reference values, The signal-to-noise ratio of the currently received signal. The target signal-to-noise ratio is preset.

3. The road surface condition detection system based on machine learning according to claim 1, characterized in that, The target discrimination unit includes a calculation subunit, a data processing subunit, and a discrimination subunit; The computing subunit is able to receive the infrared beam identifier and calculate the two-dimensional coordinates (x,y) of the interruption point in real time based on the beam space equation and timestamp. The data processing subunit can associate and cluster interruption points within a time period based on a machine learning spatiotemporal fusion network to generate a set of suspected targets. The discrimination subunit can extract features from each of the suspected targets to form an infrared feature vector. Then based on infrared feature vectors Calculate confidence score .

4. The road surface condition detection system based on machine learning according to claim 3, characterized in that, The confidence score The calculation formula is as follows: in, It is the Sigmoid activation function. This is a feature embedding function used to embed infrared feature vectors. Mapped to 64-dimensional space, These are the weighting coefficients. This is the first bias term.

5. The road surface condition detection system based on machine learning according to claim 4, characterized in that, The discrimination subunit can also make real-time judgments based on the occlusion pattern of the infrared beam.

6. The road surface condition detection system based on machine learning according to claim 1, characterized in that, The specific content of the risk simulation unit is as follows: 1) Import the real-time digital twin scene of the road segment and inject the real-time traffic flow data into the real-time digital twin scene; 2) A virtual traffic flow is generated based on a probabilistic simulation algorithm and confidence scores, and the detection interval is determined by the initial positions of all vehicles and the coordinates of the spilled material; 3) Aggregate all collision risks within the detection range and output dynamic risk values. .

7. The road surface condition detection system based on machine learning according to claim 1, characterized in that, The collision probability The calculation formula is as follows: in, For vehicles at any time Longitudinal distance from the spilled material and All are preset weighting coefficients. For action instruction function, This represents the historical collision probability.

8. The road surface condition detection system based on machine learning according to claim 1, characterized in that, The probability of secondary accidents The calculation formula is as follows: in, As the weight for accident type, Indicate whether there are other vehicles around the vehicle. This is the real-time traffic density coefficient.

9. A road surface condition detection system based on machine learning according to claim 8, characterized in that, If there are other vehicles around the vehicle, then If there are no other vehicles around the vehicle, then .

Images