General target detection method based on visual semantic prompt

Abstract

The invention relates to the technical field of driving environment target detection, and discloses a general target detection method based on visual semantic prompt, which aims at the interference of meteorological scattering and geometric distortion on bottom feature extraction in a complex urban environment, remarkably enhances the signal-to-noise ratio of a sensing system through a feature reconstruction mechanism driven by environmental perception, and improves the accuracy of target detection. Aiming at the technical pain point that a traditional image recognition model fails due to apparent heterogeneity of non-standard obstacles, physical attribute mapping from representation mode matching to physical logic judgment is completed by utilizing physical attribute decoupling and space occupation potential field modeling, dynamic feasibility boundary auditing logic based on the maximum ground attachment limit is introduced, and the dynamic feasibility boundary auditing accuracy is improved. Visual artifact interference is accurately stripped, and whether the motion state of the target is within a physical limit constraint range or not is verified through a physical axiom.

Description

Technical Field

[0001] This application relates to the field of target detection technology in driving environment, and in particular to a general target detection method based on visual semantic cues. Background Technology

[0002] With the rapid development of intelligent transportation systems, autonomous driving technology has gradually moved from road testing in limited scenarios to complex urban open road environments. In real-world high-density urban traffic scenarios, autonomous vehicles not only need to identify regular pedestrians, vehicles, and traffic signs, but also must respond in real time to a large number of highly uncertain non-standard targets and rare, unusual variables. These targets include art sculptures placed near curbs, street performers in exaggerated costumes, traffic cones overturned in the middle of the road due to sudden accidents, and various irregular industrial components accidentally spilled by freight vehicles. Such scenarios are often accompanied by complex light and shadow effects from urban buildings, strong reflection interference caused by slippery roads after rain, and dynamic occlusion caused by dense traffic flow, resulting in a high degree of heterogeneity and randomness in the visual representation of targets, constituting the extreme edge cases that autonomous driving perception systems must face.

[0003] Existing object detection technologies are mostly based on large-scale supervised learning paradigms, using deep convolutional neural networks or visual transformer structures to extract and classify visual features from massive amounts of labeled samples. While these methods perform well in recognizing predefined categories, they suffer from significant technical bottlenecks when dealing with non-standard long-tailed targets. Current detection models heavily rely on prior knowledge of texture, color, and geometric contours fixed in the training set. When the visual appearance of an object deviates from the training distribution, the visual extractor, lacking an understanding of the object's functional semantics, cannot identify its physical entity attributes, easily mistaking it for background noise or incorrect known categories. Current common improvement solutions mainly rely on long-tail data augmentation, but this approach exhibits significant lag when faced with an infinite variety of real-world variations. The core problem lies in the lack of ability to correlate the semantics and physical logic of target attributes. Existing technologies cannot utilize the object's material properties, the logic of its geometric contact points with the ground, or the consistency of its spatiotemporal trajectory to assist in identifying the presence of obstacles. This lack of semantic reasoning ability prevents the system from correcting for visual features damage or morphological distortion through semantic logic, easily leading to obstacle avoidance failures and serious traffic accidents. Summary of the Invention

[0004] This application proposes a general object detection method based on visual semantic cues to address the problems mentioned in the background art.

[0005] To achieve the above objectives, this application adopts the following technical solution: a general object detection method based on visual semantic cues, comprising the following steps:

[0006] Step S1: Obtain the original image data and real-time environmental parameters. Construct an environmental perception scaling operator using the rainfall scattering coefficient and road pitch angle in the real-time environmental parameters. Perform nonlinear activation threshold reset on the feature extraction process of the original image data to generate a net value visual feature vector. Use the net value visual feature vector to perform target pre-classification to output the visual classification probability distribution.

[0007] Step S2: Extract the functional attribute information from the net value visual feature vector generated in step S1. The functional attribute information includes the rigidity feature quantity of the target object, the consistency of edge curvature and the correlation degree of ground plane normal. Establish a spatial distribution model through the functional attribute information and calculate the spatial occupancy density distribution function to generate physical entity occupancy semantic information.

[0008] Step S3: Obtain the semantic information of the physical entity occupancy generated in step S2, combine the corresponding physical anchor point coordinates, relative velocity, road friction coefficient and gravitational acceleration constant to execute the dynamic consistency proof algorithm, calculate the dynamic compliance confidence and compare the trajectory deviation, and generate the logical proof confidence result.

[0009] Step S4: Calculate the target classification entropy value based on the visual classification probability distribution output in step S1, and perform joint gating logic judgment with the target classification entropy value and the logical counter-evidence confidence result generated in step S3; when the target classification entropy value exceeds the preset threshold, perform weight overflow compensation processing, inject the logical counter-evidence confidence result as a navigation gain factor into the detection process, guide the detection operator to perform feature enhancement on the area covered by semantic information occupied by the physical entity, and generate fusion detection results.

[0010] Furthermore, in step S1, the specific operations for constructing the environmental perception scaling operator using the rainfall scattering coefficient and road pitch angle from the real-time environmental parameters include:

[0011] First, a spatial correlation calibration is performed on the original image data, road pitch angle, and rainfall scattering coefficient through a timestamp synchronization mechanism to ensure the logical instantaneous correspondence between real-time environmental parameters and original image data;

[0012] Subsequently, using the principle of projective geometry, combined with the preset focal length parameters and installation height parameters of the visual sensor, the line-of-sight depth reconstruction operator corresponding to the coordinate position of each pixel in the original image data is determined according to the road pitch angle, in order to quantitatively characterize the actual longitudinal distance of different pixels in the physical world.

[0013] Next, based on the atmospheric transmission attenuation model, the rainfall scattering coefficient and the line-of-sight depth reconstruction operator are nonlinearly coupled to simulate the energy attenuation process of the optical signal with physical distance and scattering medium.

[0014] Finally, a Fresnel reflection correction factor is introduced to characterize the surface energy reflection properties, and the calculation results of the nonlinear coupling operation are multiplied and fused with the Fresnel reflection correction factor to finally generate an environment-aware scaling operator with non-uniform weight distribution characteristics in the image space coordinate system.

[0015] Furthermore, in step S1, the specific operations of resetting the nonlinear activation threshold and generating a net-valued visual feature vector in the feature extraction process of the original image data include:

[0016] First, the original image data is input into the feature encoding layer to extract the preliminary feature map, and the trace of the second-order gradient response tensor at each pixel coordinate position of the preliminary feature map is calculated to obtain the Laplacian curvature response value used to characterize the saliency of the local semantic signal.

[0017] Subsequently, the environment-aware scaling operator constructed in the above steps is obtained. The natural logarithm of the environment-aware scaling operator is used to perform gain adjustment on the preset standard environment activation threshold. Combined with the energy distribution characteristics of the Laplacian curvature response value, the neuron activation threshold corresponding to each pixel coordinate position is dynamically reconstructed.

[0018] In this process, when the rainfall attenuation weight represented by the environmental perception scaling operator increases, the neuron activation threshold is non-linearly increased. Environmental noise interference signals with response intensity lower than the reconstructed neuron activation threshold are filtered out through rectification logic, thereby eliminating rainfall artifacts at the bottom feature level and extracting net value visual feature vectors with physical robustness.

[0019] Finally, the net value-based visual feature vectors are connected to the classification and discrimination layer for global spatial pooling and semantic mapping processing, and the output is calculated to include the visual classification probability distribution containing the confidence vectors of the target object belonging to each preset category.

[0020] Furthermore, in step S2, the specific operations for extracting the functional attribute information from the net asset value-based visual feature vector generated in step S1 and establishing a spatial distribution model include:

[0021] First, the look-depth reconstruction operator output by step S1 is obtained, and a difference operation with respect to the spatial gradient is performed in the image coordinate system to extract the local surface normal vector used to characterize the local surface geometric orientation of the target.

[0022] Subsequently, the divergence operator operation is performed on the local surface normal vector in three-dimensional space to obtain the edge curvature consistency index used to characterize the geometric smoothness and geometric regularity of the object surface.

[0023] Next, the net value visual feature vector obtained at the current moment is extracted and compared spatiotemporally with the multi-frame historical net value visual feature vectors stored in the preset historical feature cache pool. Based on the optical flow method, the feature points of multiple frames are tracked, and the fitting residual of the feature point motion vector in the local area relative to the preset rigid body affine transformation model is calculated, thereby decoupling and generating the target object rigid feature quantity used to distinguish between rigid physical entities and non-rigid environmental interference.

[0024] Finally, the acquired local surface normal vectors, edge curvature consistency index, and target object rigidity features are mapped to a three-dimensional voxel coordinate system to establish a spatial distribution model for describing the physical properties of the object, generating functional attribute information with geometric logic and mechanical properties.

[0025] Furthermore, in step S2, the specific operations for calculating the space occupancy density distribution function and generating physical entity occupancy semantic information include:

[0026] First, the original image data output from step S1 is obtained. The average image brightness is calculated by performing statistical operations on the original image data. The material reflectivity compensation factor is calculated based on the mapping relationship between the average image brightness and the material absorption characteristics.

[0027] Subsequently, based on the road surface pitch angle output in step S1, the ground plane normal vector is determined, the orthogonal component strength of the local surface normal vector and the ground plane normal vector is calculated, and a preset ground plane offset index is introduced to perform nonlinear power operation to generate a vertical occupancy operator for quantifying the degree of vertical uplift of the local surface relative to the ground plane.

[0028] Next, the rigidity feature quantity of the target object, the edge curvature consistency index, the vertical occupancy operator, and the material reflectivity compensation factor are used as independent variables. A functional coupling operation based on potential field modeling is performed to accumulate and calculate the deterministic weight of each spatial coordinate unit occupied by the physical entity in the local integral neighborhood, so as to obtain the spatial occupancy density distribution function.

[0029] Finally, morphological clustering and boundary envelope extraction are performed on the continuous voxel regions in the spatial occupancy density distribution function that exceed the preset occupancy threshold. The output is physical entity occupancy semantic information containing three-dimensional geometric envelope coordinates and entity confidence field, which is used as input data for the subsequent execution of the dynamic consistency proof by contradiction algorithm.

[0030] Furthermore, in step S3, the specific operations for executing the dynamic consistency disproven algorithm and calculating the dynamic compliance confidence score include:

[0031] First, obtain the semantic information of physical entity occupancy output by step S2, extract the three-dimensional centroid coordinates of the continuous voxel region as the physical anchor point coordinates, and retrieve the relative speed collected in real time by the vehicle system, the road friction coefficient corrected based on the rainfall scattering coefficient output by step S1, and the preset gravitational acceleration constant.

[0032] Subsequently, using the road pitch angle output from step S1, a second-order time-domain difference operation is performed on the displacement change of the physical anchor point coordinates between adjacent sampling times to calculate the observed resultant acceleration vector of the target object at the current time, and to calculate the tangential projection modulus of the observed resultant acceleration vector on the road parallel plane.

[0033] Next, based on Coulomb's law of friction, the product of the road surface friction coefficient and the gravitational acceleration constant is calculated to obtain the maximum theoretical adhesion acceleration threshold that the physical entity can produce under the current environmental conditions.

[0034] Finally, a dynamic boundary constraint operator is introduced to calculate the over-limit ratio of the tangential projection modulus to the maximum theoretical attachment acceleration threshold. The over-limit ratio is then mapped using a nonlinear decay function to generate a dynamic compliance confidence score for quantifying whether the motion state of the target object conforms to the physical limit constraints.

[0035] Furthermore, in step S3, the specific operations for generating the logical proof of contradiction confidence results include:

[0036] First, the rigidity feature quantity of the target object and the edge curvature consistency index are obtained from the functional attribute information output by step S2. Then, a weighted mapping operation is performed on the rigidity feature quantity of the target object and the edge curvature consistency index to construct a dynamic adjustment factor for adjusting the physical logic tolerance limit.

[0037] Subsequently, the physical anchor point coordinates at the previous sampling time in step S2 are used to perform a derivation operation based on second-order state evolution to calculate the theoretical predicted position at the current sampling time. The spatial Euclidean distance between the theoretical predicted position and the physical anchor point coordinates extracted at the current sampling time in step S2 is calculated to generate a trajectory deviation vector that characterizes the degree of motion deviation.

[0038] Next, the dynamic compliance confidence level is nonlinearly scaled using a dynamic adjustment factor. By increasing the dynamic constraint weight of the region corresponding to the rigidity characteristic quantity of the target object, the differentiated physical law verification of targets with different rigidity properties is achieved.

[0039] Finally, the scaled dynamic compliance confidence score and trajectory deviation vector are multiplied and coupled logically operated to generate a logical disproven confidence score that includes the determinacy of motion trend and the determinacy of physical axiom conservation. This result is used as the judgment benchmark data for subsequent weight overflow compensation processing, providing logical disproven support based on classical mechanics axioms for the detection process.

[0040] Furthermore, in step S4, the specific operations for calculating the target classification entropy value based on the visual classification probability distribution output in step S1 and performing the joint gating logic determination include:

[0041] First, obtain the visual classification probability distribution output by step S1, and retrieve the preset category risk sensitivity factor used to characterize the obstacle avoidance risk level of different target categories. By calculating the weighted information entropy of the predicted probability of each category and the category risk sensitivity factor in the visual classification probability distribution, generate a weighted target classification entropy value used to quantify the degree of decision bottleneck in the visual perception branch.

[0042] Subsequently, the dynamic compliance confidence score in the logical proof confidence score output by step S3 is obtained. The weighted target classification entropy value and the dynamic compliance confidence score are used as input parameters of the joint gating logic. By performing nonlinear probability mapping processing based on the Boltzmann uncertainty balance equation, a joint gating signal is generated to characterize the physical property confidence score to perform hedging adjustment on visual qualitative uncertainty.

[0043] Finally, the joint gating signal is compared with the preset decision conflict threshold. If the joint gating signal exceeds the decision conflict threshold, the visual perception branch is determined to have entered the decision bottleneck state and the weight overflow compensation process is initiated. In this way, a judgment criterion for performing logical correction of visual classification ambiguity by dynamic conservation determinism is established in the perception decision chain.

[0044] Furthermore, in step S4, the specific operations for performing weight overflow compensation processing and injecting navigation gain factors include:

[0045] First, after the joint gating signal is triggered to the active state, the logical proof of contradiction confidence result and trajectory deviation vector output by step S3 are retrieved. By performing a nonlinear mapping transformation operation based on dynamic conservation determinism, a physical law deterministic weight value representing the physical conviction strength of the logical proof of contradiction confidence result is generated, and the physical law deterministic weight value is defined as the navigation gain factor.

[0046] Subsequently, the net value visual feature vector output by step S1 and the space occupancy density distribution function calculated in step S2 are obtained. The space occupancy density distribution function is used as a feature enhancement mask in the spatial dimension. The navigation gain factor is used to perform feature space remapping operator operation on the net value visual feature vector so as to realize the computational resources to perform feature reconstruction processing on the spatial region with physical entity authenticity and blurred visual representation.

[0047] Next, through the feature space remapping operator, the generated physical law deterministic weight values ​​are injected in reverse into the net value visual feature vector, so as to perform feature energy compensation on the weak visual perception area using physical axioms without changing the original image data obtained in step S1.

[0048] Furthermore, in step S4, the specific operations for generating the fusion detection result include:

[0049] First, obtain the fusion feature vector generated after the weight overflow compensation process in step S4, and simultaneously retrieve the trajectory deviation vector and the logical rebuttal confidence result output in step S3. Use the fusion feature vector, trajectory deviation vector and logical rebuttal confidence result together as input data for the multidimensional evidence consistency fusion decision.

[0050] Subsequently, using the physical geometric constraint logic in the confidence result of logical proof by contradiction, the target spatial boundary corresponding to the visual classification probability distribution output by step S1 is corrected to eliminate the spatial boundary topology error caused by visual distortion.

[0051] Next, an asymmetric coupled decision processing process, which includes classification determinism, space occupancy, and dynamic compliance, is performed to logically merge the visual classification probability distribution and the logical proof confidence result, thereby calculating and generating the final fusion detection result.

[0052] Finally, the three-dimensional physical envelope coordinates, dynamic motion trends, and category attribute labels verified by physical axioms of the target object are characterized by the fusion of detection results.

[0053] The beneficial effects of this invention are as follows:

[0054] This invention addresses the interference of meteorological scattering and geometric distortion on low-level feature extraction in complex urban environments. Through an environment-perception-driven feature reconstruction mechanism, it significantly enhances the signal-to-noise ratio of the perception system. Addressing the technical challenge of traditional image recognition models failing due to the apparent heterogeneity of non-standard obstacles, this solution utilizes physical attribute decoupling and spatial occupancy potential field modeling to complete the physical attribute mapping from appearance pattern matching to physical logic judgment. It introduces dynamic feasibility boundary audit logic based on the maximum ground adhesion limit, which not only accurately removes visual artifact interference but also verifies whether the target's motion state is within the physical limit constraints through physical axioms. Finally, based on a cross-modal compensation mechanism of visual uncertainty and physical certainty, it achieves logical compensation of feature energy under conditions of limited visual semantics, significantly improving the image recognition accuracy and robustness of multi-category targets in complex scenes. This effectively avoids the risk of missed detections caused by perceptual illusions, ensuring the security of the intelligent decision-making system. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort:

[0056] Figure 1 This is a schematic diagram of the method logic of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0058] Example

[0059] like Figure 1 As shown, this invention provides a general object detection method based on visual semantic cues, specifically including the following steps:

[0060] Step S1: Obtain the original image data and real-time environmental parameters. Construct an environmental perception scaling operator using the rainfall scattering coefficient and road pitch angle in the real-time environmental parameters. Perform nonlinear activation threshold reset on the feature extraction process of the original image data to generate a net value visual feature vector. Use the net value visual feature vector to perform target pre-classification to output the visual classification probability distribution.

[0061] In this specific embodiment, an environment-aware feature reconstruction mechanism is introduced in step S1. The activation threshold of neurons in the deep neural network is dynamically calibrated using real-time acquired external physical parameters. This enables the removal of non-entity noise features caused by meteorological scattering or geometric distortion from the underlying signal representation logic, thereby providing a consistent perception base with high signal-to-noise ratio and conforming to physical logic for subsequent semantic information mining and target entity determination processes.

[0062] Specifically, in step S1, the specific operations for constructing the environmental perception scaling operator using the rainfall scattering coefficient and road pitch angle from the real-time environmental parameters include:

[0063] First, the system performs spatial domain correlation calibration on the original image data, road pitch angle and rainfall scattering coefficient through a timestamp synchronization mechanism to ensure the logical instantaneous correspondence between real-time environmental parameters and original image data.

[0064] Specifically, the road surface pitch angle represents the instantaneous angle between the optical axis of the vision sensor and the horizontal tangent space of the road surface.

[0065] In this embodiment, the dynamic range of the road pitch angle, based on feedback from the vehicle's inertial measurement unit, is preferably set between ±0.35 radians. This value is derived from the superposition extreme value of the maximum road design gradient (such as overpass ramps or mountain city roads) under complex urban conditions and the suspension pitch angle during emergency braking. Although the gradient of conventional urban roads is relatively small, setting this wide threshold is intended to cover extreme edge conditions and the initial installation pitch error that the vision sensor may have, thereby ensuring that image projection distortion can be corrected in real time under various non-flat road conditions, providing a robust benchmark for subsequent geometric reconstruction.

[0066] The rainfall scattering coefficient represents the total scattering cross-section of rain particles on a unit length path. Its dimension is the negative first power of length. In this embodiment, the coefficient is based on the real-time infrared extinction measurement value of the rain sensor, and its value range is preferably set between 0.01 and 0.50 per meter. This value is based on the light path attenuation analysis of moderate to heavy rain levels in Mie scattering theory. Its function is to quantify the nonlinear loss of atmospheric transmittance and to guide the compensation intensity of subsequent visual features.

[0067] Subsequently, using the principle of projective geometry, combined with the preset focal length parameters and installation height parameters of the visual sensor, the line-of-sight depth reconstruction operator corresponding to the coordinate position of each pixel in the original image data is determined according to the road pitch angle, in order to quantitatively characterize the actual longitudinal distance of different pixels in the physical world.

[0068] In this embodiment, the focal length parameter of the visual sensor represents the equivalent distance from the photosensitive element to the optical center. This parameter is preferably set to 1200 pixels, and its value is derived from the camera's intrinsic parameter calibration results. Its function is to establish the mapping ratio between the pixel step size and the field of view.

[0069] The visual sensor installation height parameter represents the vertical distance between the center of the visual sensor and the road plane. In this embodiment, this parameter is preferably set to 1.65 meters. Its value is based on the physical structure design parameters of a specific vehicle model and serves to establish the vertical reference for triangulation.

[0070] The view depth reconstruction operator represents the object depth corresponding to each pixel in the image.

[0071] The specific logic is as follows: First, calculate the offset of the pixel's vertical coordinate from the center of the optical axis, and then combine the focal length parameter to obtain the viewing angle;

[0072] Secondly, the line-of-sight angle and the road surface pitch angle are summed algebraically to correct the line-of-sight vector;

[0073] Finally, the actual physical longitudinal distance is obtained by dividing the installation height parameter by the sine of the corrected backsight vector.

[0074] This view depth reconstruction operator plays a role in spatial positioning in the algorithm, solving the technical defect that image planar coordinates cannot directly reflect physical distance.

[0075] Next, based on the atmospheric transmission attenuation model, the rainfall scattering coefficient and the line-of-sight depth reconstruction operator are coupled nonlinearly to simulate the energy attenuation process of the optical signal with physical distance and scattering medium.

[0076] Specifically, an exponential decay calculation with the natural constant as the base is performed, and the negative product of the rainfall scattering coefficient and the line-of-sight depth reconstruction operator is used as the exponential term to calculate the theoretical visibility retention coefficient for each pixel position.

[0077] It is worth noting that this nonlinear coupling operation simulates the propagation logic of light waves in a non-uniform medium, and its output represents the proportion of original energy retention of visual features at different depths, which is used to calibrate local contrast collapse caused by rainfall.

[0078] Finally, a Fresnel reflection correction factor is introduced to characterize the surface energy reflection properties, and the calculation results of the nonlinear coupling operation are multiplied and fused with the Fresnel reflection correction factor to finally generate an environment-aware scaling operator with non-uniform weight distribution characteristics in the image space coordinate system.

[0079] The Fresnel reflection correction factor represents the reflection enhancement coefficient of incident light after water covers the road surface.

[0080] In this embodiment, the Fresnel reflection correction factor is preferably set between 1.05 and 1.35. Its value is based on the refractive index variation law of different roughness road surfaces under water accumulation. Its function is to logically suppress the visual artifact energy generated by the strong reflection area of ​​the road surface. By fusing with the attenuation coefficient, an environmental perception scaling operator that can simultaneously sense the depth loss and road surface reflection interference is generated.

[0081] Furthermore, in step S1, the specific operations of resetting the nonlinear activation threshold and generating a net-valued visual feature vector in the feature extraction process of the original image data include:

[0082] First, the system inputs the original image data into the feature encoding layer to extract a preliminary feature map, and calculates the trace of the second-order gradient response tensor at each pixel coordinate position of the preliminary feature map to obtain the Laplacian curvature response value used to characterize the saliency of the local semantic signal.

[0083] The preliminary feature map represents a high-dimensional feature representation of the original image after being extracted by multiple convolutional kernels, including the texture, edges, and color distribution of the target.

[0084] The Laplace curvature response value represents the curvature energy intensity of the feature space at each coordinate point.

[0085] Specifically, the second-order partial derivatives of the feature map in the horizontal and vertical directions are calculated, and the sum of the diagonal elements of the Hessian matrix formed by these derivatives is obtained.

[0086] In this embodiment, the Laplacian curvature response value is a positive scalar, and its value is based on the energy extraction logic for local salient features in digital signal processing. Its function is to identify candidate regions in the image that have physical entity attributes.

[0087] Subsequently, the environment-aware scaling operator constructed in the above steps is obtained. The natural logarithm of the environment-aware scaling operator is used to perform gain adjustment on the preset standard environment activation threshold. Combined with the energy distribution characteristics of the Laplacian curvature response value, the neuron activation threshold corresponding to each pixel coordinate position is dynamically reconstructed.

[0088] The standard environmental activation threshold represents the feature retention threshold under ideal lighting conditions.

[0089] In this embodiment, the parameter is preferably set to 0.38, and its value is based on the average statistics of a large-scale general vision task training set.

[0090] Specifically, the system performs natural logarithm operations to extract the nonlinear rate of change of the scaling operator and uses it as a scaling factor on the standard threshold. When environmental degradation is severe, the system automatically raises the activation threshold of neurons, so that only features with extremely high significance can pass through, thereby achieving dynamic signal-to-noise ratio optimization at the feature level.

[0091] In this process, when the rainfall attenuation weight represented by the environmental perception scaling operator increases, the neuron activation threshold is non-linearly increased. Environmental noise interference signals with response intensity lower than the reconstructed neuron activation threshold are filtered out by rectified logic, thereby eliminating rainfall artifacts at the underlying feature level and extracting a net value visual feature vector with physical robustness.

[0092] Specifically, a subtraction logic operation is performed to subtract the reconstructed activation threshold from the Laplacian curvature response value, and the result is truncated at zero.

[0093] This process ensures that false textures affected by rain scattering are forcibly suppressed.

[0094] Net value-based visual feature vectors represent high-purity semantic representations verified by physical laws, effectively solving the technical problem of confusion between background textures and long-tail target features.

[0095] Finally, the net value-based visual feature vectors are connected to the classification and discrimination layer for global spatial pooling and semantic mapping processing, and the output is calculated to include the visual classification probability distribution containing the confidence vectors of the target object belonging to each preset category.

[0096] This visual classification probability distribution represents the system's initial discrimination state for the target category, specifically including the probability that the target belongs to multiple categories such as pedestrians, vehicles, and traffic signs.

[0097] The output will be directly used as input for calculating the target classification entropy value in subsequent steps to assess the degree of certainty in the perception process and guide the subsequent logical decision-making process to determine whether physical logic counter-evidence compensation needs to be initiated.

[0098] It is worth noting that in long-tailed distribution scenarios of complex urban perception environments, conventional target detectors are often limited by the heterogeneity of target appearance features, such as the visual representation differences of non-standard irregular sculptures on the road or individuals with atypical clothing, or by false detection responses caused by fluctuations in the environmental background, such as mirror artifacts formed by highly reflective water on the road or reflections produced by building glass curtain walls.

[0099] To address the aforementioned perception bottleneck, this invention establishes a set of feature-based proof logic based on the axiom of physical entity exclusivity in step S2: that is, although visual representations possess morphological variability and environmental susceptibility, the occupancy logic of real physical entities in three-dimensional space possesses physical consistency and geometric insurmountability.

[0100] To this end, step S2 receives the net value visual feature vector output from step S1 as the original data base and performs non-categorical functional attribute decoupling on it, thereby upgrading the target recognition dimension from simple appearance pattern matching to physical essence attribute reasoning. By reconstructing the physical space occupancy potential field of the target in the physical dimension, deterministic recognition of non-standard obstacles at the underlying physical logic level is achieved.

[0101] Further, in step S2, the functional attribute information in the net value visual feature vector generated in step S1 is extracted. The functional attribute information includes the rigidity feature quantity of the target object, the consistency of edge curvature and the correlation degree of ground plane normal. A spatial distribution model is established through the functional attribute information and the spatial occupancy density distribution function is calculated to generate the semantic information of physical entity occupancy.

[0102] In step S2, the specific operations for extracting the functional attribute information from the net asset value-based visual feature vector generated in step S1 and establishing a spatial distribution model include:

[0103] First, the look-depth reconstruction operator output by step S1 is obtained, and a difference operation with respect to the spatial gradient is performed in the image coordinate system to extract the local surface normal vector used to characterize the local surface geometric orientation of the target.

[0104] The look-to-depth reconstruction operator represents the real physical vertical distance information corresponding to the coordinates of each pixel in the image space.

[0105] In this embodiment, the range of the view depth reconstruction operator is set between 0.5 meters and 200 meters. The value is based on the projective reconstruction calculation performed in step S1 using the focal length of the visual sensor and the installation height. Its function is to calculate the local surface normal vector that characterizes the tilt of the object's surface by providing the spatial derivative of the depth, thus providing a geometric reference for the subsequent determination of the object's three-dimensional pose.

[0106] Subsequently, the divergence operator operation is performed on the local surface normal vector in three-dimensional space to obtain the edge curvature consistency index, which is used to characterize the geometric smoothness and geometric regularity of the object surface.

[0107] The edge curvature uniformity index represents the rate of change of the surface normal vector field of the target object.

[0108] In this embodiment, the edge curvature consistency index is preferably set in the real number range between 0.0 and 10.0. Its value is based on the structural description of the second-order manifold of the curvature tensor in differential geometry. It is used to quantify the orderliness of the surface geometry. Its function is to distinguish between rigid solid targets with regular geometric contours and environmental noise edges that present random pulse characteristics, thereby eliminating visual interference that does not have physical entity attributes in the geometric feature dimension.

[0109] Next, the net value visual feature vector obtained at the current moment is extracted and spatiotemporally compared with the multi-frame historical net value visual feature vectors stored in the preset historical feature cache pool. Based on the optical flow method, the feature points of multiple frames are tracked, and the fitting residual of the feature point motion vector in the local area relative to the preset rigid body affine transformation model is calculated, thereby decoupling and generating the target object rigid feature quantity used to distinguish between rigid physical entities and non-rigid environmental interference.

[0110] Specifically, when a rigid object is in motion, the optical flow vectors of each feature point on its surface should satisfy the same homography matrix or affine transformation constraint. If the fitting residual of a certain region (such as water ripples or smoke) exceeds a preset threshold, it indicates that the motion of its internal feature points does not have rigid body consistency, thus determining that its rigidity feature quantity is low.

[0111] The rigidity of a target represents its ability to maintain its internal topological structure during movement or changes in viewpoint.

[0112] In this embodiment, the value range of the rigidity feature of the target object is set between 0.0 and 1.0. Its value is based on the consistency measure of the Hilbert-Schmidt norm of feature points between adjacent video frames. It plays the role of adjusting the activation weight in the algorithm. By giving higher confidence to high-rigidity regions, it effectively filters out non-rigid false feature responses caused by light and shadow shaking, smoke or water refraction.

[0113] Finally, the acquired local surface normal vectors, edge curvature consistency index, and target object rigidity features are mapped to a three-dimensional voxel coordinate system to establish a spatial distribution model for describing the physical properties of the object, generating functional attribute information with geometric logic and mechanical properties.

[0114] This step achieves the transformation from pixel patterns to physical entity attribute descriptions by aggregating discrete attribute features into a feature field with spatial continuity.

[0115] In step S2, the specific operations for calculating the spatial occupancy density distribution function and generating physical entity occupancy semantic information include:

[0116] First, the original image data output from step S1 is obtained. The average image brightness is calculated by performing statistical operations on the original image data. Then, the material reflectivity compensation factor is calculated based on the mapping relationship between the average image brightness and the material absorption characteristics.

[0117] The material reflectivity compensation factor represents the intensity of the system's gain adjustment on visual energy for materials with different physical properties.

[0118] In this embodiment, the value range of the material reflectivity compensation factor is preferably set between 0.8 and 1.5. The value is based on typical road targets such as light-absorbing materials like asphalt and reflective materials like metal, and their corresponding grayscale statistical models. Its function is to automatically compensate for the feature activation energy through the material reflectivity compensation factor when the surface of an object is detected to be a low reflectivity material, thereby ensuring the feature stability of the physical entity under complex lighting backgrounds.

[0119] Subsequently, based on the road surface pitch angle output in step S1, the ground plane normal vector is determined, the orthogonal component strength of the local surface normal vector and the ground plane normal vector is calculated, and a preset ground plane offset index is introduced to perform nonlinear power operation to generate a vertical occupancy operator for quantifying the degree of vertical uplift of the local surface relative to the ground plane.

[0120] The road surface pitch angle, as the real-time angular offset between the vehicle coordinate system and the road plane, directly defines the projective projection of the gravity direction in the image, and thus determines the ground plane normal vector.

[0121] The ground plane offset index represents the physical suppression sensitivity to non-vertical plane interference. In this embodiment, the ground plane offset index is preferably set to three points zero. Its value is based on the probability statistics of the vertical boundary distribution of obstacles in urban environments. Its function is to exclude the planar features and ground textures of the ground through nonlinear exponential suppression logic, so that the vertical occupancy operator only produces a high-intensity response to physical entities with vertical support features.

[0122] Next, the rigidity feature of the target object, the edge curvature consistency index, the vertical occupancy operator, and the material reflectivity compensation factor are used as independent variables. A functional coupling operation based on potential field modeling is performed to accumulate and calculate the deterministic weight of each spatial coordinate unit occupied by the physical entity in the local integral neighborhood, thereby obtaining the spatial occupancy density distribution function.

[0123] The space occupancy density distribution function represents the probability distribution intensity of each voxel unit in space being actually occupied by a physical entity.

[0124] By performing multiplication and integration operations on cross-modal features, mechanics, geometry, and environmental compensation logic are deeply coupled to form a spatial entity determination field that does not depend on specific classification labels.

[0125] Finally, morphological clustering and boundary envelope extraction are performed on the continuous voxel regions in the spatial occupancy density distribution function that exceed the preset occupancy threshold. The output is physical entity occupancy semantic information containing three-dimensional geometric envelope coordinates and entity confidence field, which is used as input data for the subsequent execution of the dynamic consistency proof by contradiction algorithm.

[0126] The preset occupancy threshold represents the confidence boundary for determining whether a physical entity exists in the three-dimensional voxel space. In this embodiment, the preset occupancy threshold is preferably set to 0.65. Its value is based on the outlier analysis of the high-frequency noise modulus of the visual sensor. Its function is to filter out isolated shot noise in the space and ensure that the output physical entity occupancy semantic information can accurately describe the spatial location, volume envelope, and confidence of the obstacle as a real physical entity.

[0127] It is worth noting that although step S2 establishes the semantic information of physical entity occupancy with spatial distribution determinism by decoupling the target functional attributes in multiple dimensions, it is still difficult to completely remove non-physical interference signals with high physical deception based solely on the spatial occupancy representation of discrete single frames under the extreme complex perception conditions in urban areas.

[0128] For example, high-fidelity road surface total reflection image artifacts, or static pixel redundancy caused by damage to the surface of visual sensor hardware, such interference often exhibit camouflage rigidity characteristics that are highly similar to real obstacles in spatial topology and geometry.

[0129] To address the aforementioned technical bottlenecks, this invention introduces a logic auditing mechanism based on temporal dynamic constraints in step S3. By utilizing the conservation laws of classical mechanics, it performs adversarial logical proof against the rationality of the movement of physical entities occupying semantic information.

[0130] Further, in step S3, the semantic information of the physical entity occupancy generated in step S2 is obtained, and the dynamic consistency proof algorithm is executed in combination with the corresponding physical anchor point coordinates, relative velocity, road surface friction coefficient and gravitational acceleration constant. The dynamic compliance confidence score is calculated and the trajectory deviation is compared to generate the logical proof confidence score result.

[0131] In step S3, the specific operations for executing the dynamic consistency disproven algorithm and calculating the dynamic compliance confidence score include:

[0132] First, obtain the semantic information of physical entity occupancy output by step S2, extract the three-dimensional centroid coordinates of the continuous voxel region as the physical anchor point coordinates, and retrieve the relative speed collected in real time by the vehicle system, the road friction coefficient corrected based on the rainfall scattering coefficient output by step S1, and the preset gravitational acceleration constant.

[0133] The physical anchor point coordinates represent the center point of motion of the detected entity in three-dimensional space, and their coordinate values ​​are distributed within the effective sensing range of the sensor.

[0134] The coordinates are derived from the centroids of the continuous connected domains corresponding to the physical entity occupancy density distribution function extracted in step S2. Their purpose is to provide spatial anchor points for subsequent motion trajectory monitoring and acceleration calculation.

[0135] Relative speed represents the instantaneous velocity of an object relative to the vehicle.

[0136] The road surface friction coefficient represents the maximum adhesion potential provided by the road surface. In this embodiment, this parameter is preferably set between 0.15 and 0.95, and its value is based on the inversion of the road surface water film thickness by the rainfall scattering coefficient obtained in step S1.

[0137] The gravitational acceleration constant is taken as a standard physical constant (approximately 9.8 meters per square second), and its function is to determine the physical grip limit in conjunction with the friction coefficient.

[0138] Subsequently, using the road pitch angle output from step S1, a second-order time-domain difference operation is performed on the displacement change of the physical anchor point coordinates between adjacent sampling times to calculate the observed resultant acceleration vector of the target object at the current time, and to calculate the tangential projection modulus of the observed resultant acceleration vector on the road surface parallel plane.

[0139] Specifically, by performing discrete differentiation on the physical anchor point coordinates of three consecutive frames, the instantaneous acceleration of the target in the sensor coordinate system is calculated.

[0140] Considering the influence of road slope on dynamic decomposition, a rotation matrix is ​​constructed using the road pitch angle, and the observed resultant acceleration vector is projected onto the road tangential plane, thereby removing the influence of gravity component and retaining only the tangential projection modulus used to drive the planar motion of the object.

[0141] This process aims to obtain the actual maneuverability exhibited by the target.

[0142] Next, based on Coulomb's law of friction, the product of the road surface friction coefficient and the gravitational acceleration constant is calculated to obtain the maximum theoretical adhesion acceleration threshold that a physical entity can generate under the current environmental conditions. The maximum theoretical adhesion acceleration threshold represents the upper limit of acceleration that any physical entity (whether a vehicle tire or a pedestrian's shoe sole) that relies on ground friction to accelerate, decelerate, or turn under the current road surface slipperiness.

[0143] This threshold forms the boundary of the friction circle for distinguishing between real and false targets. For example, on a slippery road surface, a real physical entity cannot generate lateral acceleration that exceeds the friction limit, while visual artifacts (such as light spot movement or mirror reflections) are not subject to this physical law.

[0144] Finally, a dynamic boundary constraint operator is introduced to calculate the over-limit ratio of the tangential projection modulus to the maximum theoretical attachment acceleration threshold. The over-limit ratio is then mapped using a nonlinear decay function to generate a dynamic compliance confidence score for quantifying whether the motion state of the target object conforms to the physical limit constraints.

[0145] Specifically, the quotient of the tangential projection modulus divided by the maximum theoretical attachment acceleration threshold is calculated.

[0146] When the quotient is less than or equal to one, it indicates that the target motion is within the physical limits and the dynamic compliance confidence level approaches one. When the quotient is significantly greater than one (i.e., the observed acceleration far exceeds the physical limits), the excess part is penalized using a sigmoid-form decay function, causing the output dynamic compliance confidence level to drop rapidly.

[0147] The confidence score will be directly used to identify and suppress false targets that, although visually textured, exhibit kinematic shifts, unnatural high-frequency jitter, or violate tribomechanical constraints (such as water surface reflections, virtual images of glass curtain walls, etc.), thereby ensuring the physical authenticity of the detection results.

[0148] In step S3, the specific operations for generating the logical proof of contradiction confidence results include:

[0149] First, the rigidity feature quantity of the target object and the edge curvature consistency index are obtained from the functional attribute information output by step S2. Then, a weighted mapping operation is performed on the rigidity feature quantity of the target object and the edge curvature consistency index to construct a dynamic adjustment factor for adjusting the physical logic tolerance limit.

[0150] The dynamic adjustment factor represents the system's fault tolerance redundancy when performing energy consistency verification. In this embodiment, the dynamic adjustment factor is preferably set between 0.1 and 3.0.

[0151] The value of the dynamic adjustment factor is based on the consistency of the target geometric features extracted in step S2:

[0152] When the target object has a higher rigidity and a more regular edge curvature distribution, the smaller the value of the dynamic adjustment factor, the more stringent the dynamic constraints of the system on the entity.

[0153] The role of the dynamic adjustment factor is to enable the classification and auditing of targets with different physical properties, and to enhance the sensitivity of the sensing system to detect non-standard rigid entities.

[0154] Subsequently, the physical anchor coordinates from the previous sampling time in step S2 are used to perform a derivation operation based on second-order state evolution to calculate the theoretical predicted position at the current sampling time. The spatial Euclidean distance between the theoretical predicted position and the physical anchor coordinates extracted at the current sampling time in step S2 is also calculated to generate a trajectory deviation vector characterizing the degree of motion deviation.

[0155] Specifically, a Taylor series expansion including an acceleration factor is used to extrapolate forward based on the displacement and velocity states of previous time steps.

[0156] The trajectory deviation vector represents the degree of deviation between the measured physical trajectory and the theoretical dynamic trajectory. In this embodiment, its typical value range is from 0.1 meters to 5 meters. The trajectory deviation vector is obtained by subtraction of three-dimensional spatial geometric coordinates and is used to measure the physical compatibility of the target object's motion trend.

[0157] Next, the dynamic compliance confidence level is nonlinearly scaled using a dynamic adjustment factor. By increasing the dynamic constraint weight of the region corresponding to the rigidity characteristic quantity of the target object, the differentiated physical law verification of targets with different rigidity properties is achieved.

[0158] Specifically, the division logic is executed, and the dynamic out-of-limit residual term is sensitively scaled by a dynamic adjustment factor before being input into the mapping function.

[0159] This mechanism causes the system to drastically reduce the score for minor dynamic violations of rigid targets, thereby distinguishing real rigid obstacles from visual projection artifacts that move synchronously with the vehicle at the underlying logic level.

[0160] Finally, the scaled dynamic compliance confidence score and trajectory deviation vector are multiplied and coupled logically operated to generate a logical disproven confidence score that includes the determinacy of motion trend and the determinacy of physical axiom conservation. This result is used as the judgment benchmark data for subsequent weight overflow compensation processing, providing logical disproven support based on classical mechanics axioms for the detection process.

[0161] The confidence result of logical proof by contradiction represents a multi-dimensional judgment on the authenticity of the target entity.

[0162] By fusing instantaneous energy balance audits with long-term trajectory prediction biases, this result can establish the legitimacy of an object in the three-dimensional physical world without relying on prior classification labels.

[0163] The result will be fully passed to step S4, serving as the core guiding vector that guides the detection operator to achieve accurate focusing in complex environments.

[0164] It is worth noting that after completing the visual feature net value reconstruction based on step S1, the non-categorical physical entity occupancy determination based on step S2, and the unit mass dynamic conservation counter-evidence audit based on step S3, the entire perception process enters the final decision fusion stage of the multi-dimensional evidence chain.

[0165] Specifically, under extreme autonomous driving perception conditions, due to the significant apparent heterogeneity of long-tail targets, such as road obstacles with non-standard shapes or individuals that exhibit visual distortion under specific lighting conditions, visual classifiers often experience classification hesitation at the decision-making level.

[0166] This phenomenon manifests at the data feature level as a high-entropy state in the visual classification probability distribution, making it difficult for traditional filtering logic based on fixed hard thresholds to suppress false alarms while avoiding missed detections.

[0167] In step S4, the present invention establishes a cross-modal evidence chain redemption mechanism. When the uncertainty of the visual branch at the qualitative level reaches the preset technical boundary, the system no longer relies solely on visual representation, but extracts the logical counter-evidence confidence result with physical logical certainty output by step S3, and performs weight overflow compensation processing on the visual channel.

[0168] Specifically, by transforming the counter-evidence information from the dynamic dimension into a navigation gain at the semantic level, the detection operator is forced to perform feature enhancement and refocusing on the specific spatial voxel region covered by the semantic information occupied by the physical entity in step S2. This achieves a deep fusion of visual discrimination, spatial geometry, and dynamic axioms in the final decision-making stage.

[0169] Furthermore, in step S4, the target classification entropy value is calculated based on the visual classification probability distribution output in step S1, and the target classification entropy value is combined with the logical counter-evidence confidence result generated in step S3 for joint gating logic judgment; when the target classification entropy value exceeds the preset threshold, weight overflow compensation processing is performed, and the logical counter-evidence confidence result is injected into the detection process as a navigation gain factor to guide the detection operator to perform feature enhancement on the area covered by semantic information occupied by the physical entity, and generate fusion detection results.

[0170] In step S4, the specific operations of calculating the target classification entropy value based on the visual classification probability distribution output in step S1 and performing joint gating logic determination include:

[0171] First, obtain the visual classification probability distribution output by step S1, and retrieve the preset category risk sensitivity factor used to characterize the obstacle avoidance risk level of different target categories. By calculating the weighted information entropy of the predicted probability of each category in the visual classification probability distribution and the category risk sensitivity factor, a weighted target classification entropy value used to quantify the degree of confusion in the visual perception branch judgment is generated.

[0172] Specifically, the cumulative operation guided by the negative sign is performed to calculate the product of the predicted probability of each target category in the visual classification probability distribution and the binary logarithm of that probability, and then multiply it by the corresponding category risk sensitivity factor.

[0173] The category risk sensitivity factor represents the semantic priority weight of different traffic participants in terms of obstacle avoidance safety.

[0174] In this embodiment, for high-risk categories such as pedestrians or non-motorized vehicles, the category risk sensitivity factor is preferably set to 2.5; for static environmental targets such as road signs or buildings, the category risk sensitivity factor is set to 1.0. The value is based on the safety risk classification model in the autonomous driving obstacle avoidance specification. Its function is to artificially amplify the uncertainty response of the visual perception branch to the judgment of high-risk targets, thereby generating a weighted target classification entropy value.

[0175] The weighted target classification entropy value represents the strength of the logical conflict in the current perception decision chain regarding the qualitative nature of the target. Its dimension is bits, and its typical value range is set between zero bits and 3.2 bits. Its value is based on the nonlinear measure of the probability distribution of each category. Its adjustment role in the algorithm is to accurately identify whether the visual perception channel has reached the bottleneck boundary of technical decision-making.

[0176] Subsequently, the system obtains the dynamic compliance confidence score from the logical proof confidence score output by step S3, uses the weighted target classification entropy value and the dynamic compliance confidence score as input parameters of the joint gating logic, and calculates and generates a joint gating signal to characterize the hedging adjustment of physical property confidence against visual qualitative uncertainty by performing nonlinear probability mapping processing based on the Boltzmann uncertainty balance equation.

[0177] Specifically, an exponential mapping function is constructed, which is the difference between the product of the weighted target classification entropy value and the dynamic compliance confidence value and the decision conflict threshold, divided by the system decision temperature coefficient, and then used to calculate the negative exponential term of the natural constant.

[0178] The decision conflict threshold represents the energy boundary for the system to trigger physical logic compensation. In this embodiment, the decision conflict threshold is preferably set to 1.2 bits. Its value is based on the classification entropy distribution quantile of the visual encoder on the standard verification set. Its function is to define the judgment criterion for initiating the weight overflow compensation process.

[0179] The system determines that the temperature coefficient represents the smoothness of the transition between visual hesitation and physical certainty in the decision-making logic. It is a dimensionless constant. In this embodiment, the coefficient is preferably set to 0.15. Its value is based on the system's sensitivity calibration to sensor noise interference and is used to adjust the sensitivity of the gating signal to fluctuations in the input variable.

[0180] The final generated joint gating signal has a value range between zero and one, representing the strength of the logical preference for the overflow of physical deterministic energy and the intervention of visual feature enhancement.

[0181] Finally, the joint gating signal is compared with the preset decision conflict threshold. If the joint gating signal exceeds the decision conflict threshold, the visual perception branch is determined to have entered the decision bottleneck state and the weight overflow compensation process is initiated.

[0182] This process establishes a criterion in the perception-decision chain for logical compensation of visual classification ambiguity based on dynamic conservation determinism.

[0183] In step S4, the specific operations for performing weight overflow compensation processing and injecting navigation gain factors include:

[0184] First, after the joint gating signal is triggered into the active state, the logical proof of contradiction confidence result and trajectory deviation vector output by step S3 are retrieved. By performing a nonlinear mapping transformation operation based on dynamic conservation determinism, a physical law deterministic weight value representing the physical conviction strength of the logical proof of contradiction confidence result is generated, and the physical law deterministic weight value is defined as the navigation gain factor.

[0185] The navigation gain factor represents the guiding weight of the physical confidence generated based on the laws of dynamics on the visual feature space. In this embodiment, the navigation gain factor is preferably set between 1.0 and 5.0. Its value is based on the inverse correlation mapping between the confidence of dynamic compliance and the magnitude of trajectory deviation. Its function is to transform the certainty of physical laws into a guiding weight operator, guiding the detection operator to refocus.

[0186] Subsequently, the net value-based visual feature vector output from step S1 and the spatial occupancy density distribution function calculated in step S2 are obtained. The spatial occupancy density distribution function is used as a feature enhancement mask in the spatial dimension, and the feature space remapping operator operation is performed on the net value-based visual feature vector using the navigation gain factor.

[0187] Specifically, element-wise multiplication is performed to combine the space occupancy density distribution function with the navigation gain factor.

[0188] The spatial occupancy density distribution function represents the deterministic weight field of spatial coordinate units occupied by physical entities. Its value is based on the physical attribute coupling result extracted in step S2, and its function is to accurately define the spatial geometric boundary that needs to be enhanced.

[0189] Next, through the feature space remapping operator, the generated physical law deterministic weight values ​​are injected in reverse into the net value visual feature vector, so as to perform feature energy compensation on the weak visual perception area using physical axioms without changing the original image data obtained in step S1.

[0190] The fused feature vectors characterize the composite perceptual tensor enhanced by dynamic confidence;

[0191] This composite perceptual tensor performs logical compensation on semantically degraded visual features through deterministic measures of physical laws, overcoming the perceptual limitations caused by the lack of apparent texture features, thereby achieving target spatial localization based on the physical entity occupancy attribute.

[0192] In step S4, the specific operations for generating the fusion detection results include:

[0193] First, the system obtains the fusion feature vector generated after the weight overflow compensation process in step S4, and simultaneously retrieves the trajectory deviation vector and the logical rebuttal confidence result output in step S3. The fusion feature vector, trajectory deviation vector and logical rebuttal confidence result are used together as input data for the multidimensional evidence consistency fusion decision.

[0194] Subsequently, using the physical geometric constraint logic in the confidence result of logical proof by contradiction, the target spatial boundary corresponding to the visual classification probability distribution output by step S1 is corrected to eliminate the spatial boundary topology error caused by visual distortion.

[0195] Specifically, the system performs a consistent fit between the geometric boundary and the predicted dynamic trajectory, using the motion continuity of rigid physical entities to constrain the irrational expansion of the visual boundary.

[0196] Next, an asymmetric coupled decision processing process, which includes classification determinism, space occupancy, and dynamic compliance, is performed to logically merge the visual classification probability distribution and the logical counter-evidence confidence results, thereby calculating and generating the final fusion detection result.

[0197] Specifically, the decision weights of the visual branch and the physical counter-evidence branch are dynamically allocated based on the weighted target classification entropy value, so that physical logic can perform compensatory confirmation when visual hesitation occurs.

[0198] Finally, the three-dimensional physical envelope coordinates, dynamic motion trends, and category attribute labels verified by physical axioms of the target object are characterized by the fusion of detection results.

[0199] It is worth noting that the fusion detection results also include collision risk priority.

[0200] Collision risk priority represents the potential threat level of a target entity to the vehicle's path of travel.

[0201] In this embodiment, the parameter is preferably set to level one to ten, where level ten represents extremely high danger. Its value is based on the geometric overlap probability between the relative velocity vector of the target object and the predicted path of the vehicle, and is corrected in combination with the dynamic compliance confidence level. Its function is to provide the vehicle planning and control layer with obstacle avoidance triggering instructions with physical consistency guarantees.

[0202] Through the above process, this invention achieves highly robust detection results by utilizing the deep coupling of visual cues and physical axioms in extreme long-tail scenarios.

[0203] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A general object detection method based on visual semantic cues, characterized in that, Includes the following steps: Step S1: Obtain the original image data and real-time environmental parameters. Construct an environmental perception scaling operator using the rainfall scattering coefficient and road pitch angle in the real-time environmental parameters. Perform nonlinear activation threshold reset on the feature extraction process of the original image data to generate a net value visual feature vector. Use the net value visual feature vector to perform target pre-classification to output the visual classification probability distribution. Step S2: Extract the functional attribute information from the net value visual feature vector generated in step S1. The functional attribute information includes the rigidity feature quantity of the target object, the consistency of edge curvature and the correlation degree of ground plane normal. Establish a spatial distribution model through the functional attribute information and calculate the spatial occupancy density distribution function to generate physical entity occupancy semantic information. Step S3: Obtain the semantic information of the physical entity occupancy generated in step S2, combine the corresponding physical anchor point coordinates, relative velocity, road friction coefficient and gravitational acceleration constant to execute the dynamic consistency proof algorithm, calculate the dynamic compliance confidence and compare the trajectory deviation, and generate the logical proof confidence result. Step S4: Calculate the target classification entropy value based on the visual classification probability distribution output in step S1, and perform joint gating logic judgment with the target classification entropy value and the logical proof confidence result generated in step S3; When the target classification entropy value exceeds the preset threshold, weight overflow compensation processing is performed. The logical proof confidence result is injected into the detection process as a navigation gain factor to guide the detection operator to perform feature enhancement on the area covered by semantic information occupied by the physical entity, and generate fusion detection results.

2. The general object detection method based on visual semantic cues according to claim 1, characterized in that, In step S1, the specific operations for constructing the environmental perception scaling operator using the rainfall scattering coefficient and road pitch angle from real-time environmental parameters include: First, a spatial correlation calibration is performed on the original image data, road pitch angle, and rainfall scattering coefficient through a timestamp synchronization mechanism to ensure the logical instantaneous correspondence between real-time environmental parameters and original image data; Subsequently, using the principle of projective geometry, combined with the preset focal length parameters and installation height parameters of the visual sensor, the line-of-sight depth reconstruction operator corresponding to the coordinate position of each pixel in the original image data is determined according to the road pitch angle, in order to quantitatively characterize the actual longitudinal distance of different pixels in the physical world. Next, based on the atmospheric transmission attenuation model, the rainfall scattering coefficient and the line-of-sight depth reconstruction operator are nonlinearly coupled to simulate the energy attenuation process of the optical signal with physical distance and scattering medium. Finally, a Fresnel reflection correction factor is introduced to characterize the surface energy reflection properties, and the calculation results of the nonlinear coupling operation are multiplied and fused with the Fresnel reflection correction factor to finally generate an environment-aware scaling operator with non-uniform weight distribution characteristics in the image space coordinate system.

3. The general object detection method based on visual semantic cues according to claim 2, characterized in that, In step S1, the specific operations of resetting the nonlinear activation threshold and generating a net value visual feature vector in the feature extraction process of the original image data include: First, the original image data is input into the feature encoding layer to extract the preliminary feature map, and the trace of the second-order gradient response tensor at each pixel coordinate position of the preliminary feature map is calculated to obtain the Laplacian curvature response value used to characterize the saliency of the local semantic signal. Subsequently, the environment-aware scaling operator constructed in the above steps is obtained. The natural logarithm of the environment-aware scaling operator is used to perform gain adjustment on the preset standard environment activation threshold. Combined with the energy distribution characteristics of the Laplacian curvature response value, the neuron activation threshold corresponding to each pixel coordinate position is dynamically reconstructed. In this process, when the rainfall attenuation weight represented by the environmental perception scaling operator increases, the neuron activation threshold is non-linearly increased. Environmental noise interference signals with response intensity lower than the reconstructed neuron activation threshold are filtered out through rectification logic, thereby eliminating rainfall artifacts at the bottom feature level and extracting net value visual feature vectors with physical robustness. Finally, the net value-based visual feature vectors are connected to the classification and discrimination layer for global spatial pooling and semantic mapping processing, and the output is calculated to include the visual classification probability distribution containing the confidence vectors of the target object belonging to each preset category.

4. The general object detection method based on visual semantic cues according to claim 3, characterized in that, In step S2, the specific operations for extracting the functional attribute information from the net asset value-based visual feature vector generated in step S1 and establishing a spatial distribution model include: First, the look-depth reconstruction operator output by step S1 is obtained, and a difference operation with respect to the spatial gradient is performed in the image coordinate system to extract the local surface normal vector used to characterize the local surface geometric orientation of the target. Subsequently, the divergence operator operation is performed on the local surface normal vector in three-dimensional space to obtain the edge curvature consistency index used to characterize the geometric smoothness and geometric regularity of the object surface. Next, the net value visual feature vector obtained at the current moment is extracted and compared spatiotemporally with the multi-frame historical net value visual feature vectors stored in the preset historical feature cache pool. Based on the optical flow method, the feature points of multiple frames are tracked, and the fitting residual of the feature point motion vector in the local area relative to the preset rigid body affine transformation model is calculated, thereby decoupling and generating the target object rigid feature quantity used to distinguish between rigid physical entities and non-rigid environmental interference. Finally, the acquired local surface normal vectors, edge curvature consistency index, and target object rigidity features are mapped to a three-dimensional voxel coordinate system to establish a spatial distribution model for describing the physical properties of the object, generating functional attribute information with geometric logic and mechanical properties.

5. The general object detection method based on visual semantic cues according to claim 4, characterized in that, In step S2, the specific operations for calculating the spatial occupancy density distribution function and generating physical entity occupancy semantic information include: First, the original image data output from step S1 is obtained. The average image brightness is calculated by performing statistical operations on the original image data. The material reflectivity compensation factor is calculated based on the mapping relationship between the average image brightness and the material absorption characteristics. Subsequently, based on the road surface pitch angle output in step S1, the ground plane normal vector is determined, the orthogonal component strength of the local surface normal vector and the ground plane normal vector is calculated, and a preset ground plane offset index is introduced to perform nonlinear power operation to generate a vertical occupancy operator for quantifying the degree of vertical uplift of the local surface relative to the ground plane. Next, the rigidity feature quantity of the target object, the edge curvature consistency index, the vertical occupancy operator, and the material reflectivity compensation factor are used as independent variables. A functional coupling operation based on potential field modeling is performed to accumulate and calculate the deterministic weight of each spatial coordinate unit occupied by the physical entity in the local integral neighborhood, so as to obtain the spatial occupancy density distribution function. Finally, morphological clustering and boundary envelope extraction are performed on the continuous voxel regions in the spatial occupancy density distribution function that exceed the preset occupancy threshold. The output is physical entity occupancy semantic information containing three-dimensional geometric envelope coordinates and entity confidence field, which is used as input data for the subsequent execution of the dynamic consistency proof by contradiction algorithm.

6. The general object detection method based on visual semantic cues according to claim 5, characterized in that, In step S3, the specific operations for executing the dynamic consistency disproven algorithm and calculating the dynamic compliance confidence score include: First, obtain the semantic information of physical entity occupancy output by step S2, extract the three-dimensional centroid coordinates of the continuous voxel region as the physical anchor point coordinates, and retrieve the relative speed collected in real time by the vehicle system, the road friction coefficient corrected based on the rainfall scattering coefficient output by step S1, and the preset gravitational acceleration constant. Subsequently, using the road pitch angle output from step S1, a second-order time-domain difference operation is performed on the displacement change of the physical anchor point coordinates between adjacent sampling times to calculate the observed resultant acceleration vector of the target object at the current time, and to calculate the tangential projection modulus of the observed resultant acceleration vector on the road parallel plane. Next, based on Coulomb's law of friction, the product of the road surface friction coefficient and the gravitational acceleration constant is calculated to obtain the maximum theoretical adhesion acceleration threshold that the physical entity can produce under the current environmental conditions. Finally, a dynamic boundary constraint operator is introduced to calculate the over-limit ratio of the tangential projection modulus to the maximum theoretical attachment acceleration threshold. The over-limit ratio is then mapped using a nonlinear decay function to generate a dynamic compliance confidence score for quantifying whether the motion state of the target object conforms to the physical limit constraints.

7. The general object detection method based on visual semantic cues according to claim 6, characterized in that, In step S3, the specific operations for generating the logical proof of contradiction confidence results include: First, the rigidity feature quantity of the target object and the edge curvature consistency index are obtained from the functional attribute information output by step S2. Then, a weighted mapping operation is performed on the rigidity feature quantity of the target object and the edge curvature consistency index to construct a dynamic adjustment factor for adjusting the physical logic tolerance limit. Subsequently, the physical anchor point coordinates at the previous sampling time in step S2 are used to perform a derivation operation based on second-order state evolution to calculate the theoretical predicted position at the current sampling time. The spatial Euclidean distance between the theoretical predicted position and the physical anchor point coordinates extracted at the current sampling time in step S2 is calculated to generate a trajectory deviation vector that characterizes the degree of motion deviation. Next, the dynamic compliance confidence level is nonlinearly scaled using a dynamic adjustment factor. By increasing the dynamic constraint weight of the region corresponding to the rigidity characteristic quantity of the target object, the differentiated physical law verification of targets with different rigidity properties is achieved. Finally, the scaled dynamic compliance confidence score and trajectory deviation vector are multiplied and coupled logically operated to generate a logical disproven confidence score that includes the determinacy of motion trend and the determinacy of physical axiom conservation. This result is used as the judgment benchmark data for subsequent weight overflow compensation processing, providing logical disproven support based on classical mechanics axioms for the detection process.

8. The general object detection method based on visual semantic cues according to claim 7, characterized in that, In step S4, the specific operations of calculating the target classification entropy value based on the visual classification probability distribution output in step S1 and performing joint gating logic determination include: First, obtain the visual classification probability distribution output by step S1, and retrieve the preset category risk sensitivity factor used to characterize the obstacle avoidance risk level of different target categories. By calculating the weighted information entropy of the predicted probability of each category and the category risk sensitivity factor in the visual classification probability distribution, generate a weighted target classification entropy value used to quantify the degree of decision bottleneck in the visual perception branch. Subsequently, the dynamic compliance confidence score in the logical proof confidence score output by step S3 is obtained. The weighted target classification entropy value and the dynamic compliance confidence score are used as input parameters of the joint gating logic. By performing nonlinear probability mapping processing based on the Boltzmann uncertainty balance equation, a joint gating signal is generated to characterize the physical property confidence score to perform hedging adjustment on visual qualitative uncertainty. Finally, the joint gating signal is compared with the preset decision conflict threshold. If the joint gating signal exceeds the decision conflict threshold, the visual perception branch is determined to have entered the decision bottleneck state and the weight overflow compensation process is initiated. In this way, a judgment criterion for performing logical correction of visual classification ambiguity by dynamic conservation determinism is established in the perception decision chain.

9. The general object detection method based on visual semantic cues according to claim 8, characterized in that, In step S4, the specific operations for performing weight overflow compensation processing and injecting navigation gain factors include: First, after the joint gating signal is triggered to the active state, the logical proof of contradiction confidence result and trajectory deviation vector output by step S3 are retrieved. By performing a nonlinear mapping transformation operation based on dynamic conservation determinism, a physical law deterministic weight value representing the physical conviction strength of the logical proof of contradiction confidence result is generated, and the physical law deterministic weight value is defined as the navigation gain factor. Subsequently, the net value visual feature vector output by step S1 and the space occupancy density distribution function calculated in step S2 are obtained. The space occupancy density distribution function is used as a feature enhancement mask in the spatial dimension. The navigation gain factor is used to perform feature space remapping operator operation on the net value visual feature vector so as to realize the computational resources to perform feature reconstruction processing on the spatial region with physical entity authenticity and blurred visual representation. Next, through the feature space remapping operator, the generated physical law deterministic weight values ​​are injected in reverse into the net value visual feature vector, so as to perform feature energy compensation on the weak visual perception area using physical axioms without changing the original image data obtained in step S1.

10. The general object detection method based on visual semantic cues according to claim 9, characterized in that, In step S4, the specific operations for generating the fusion detection results include: First, obtain the fusion feature vector generated after the weight overflow compensation process in step S4, and simultaneously retrieve the trajectory deviation vector and the logical rebuttal confidence result output in step S3. Use the fusion feature vector, trajectory deviation vector and logical rebuttal confidence result together as input data for the multidimensional evidence consistency fusion decision. Subsequently, using the physical geometric constraint logic in the confidence result of logical proof by contradiction, the target spatial boundary corresponding to the visual classification probability distribution output by step S1 is corrected to eliminate the spatial boundary topology error caused by visual distortion. Next, an asymmetric coupled decision processing process, which includes classification determinism, space occupancy, and dynamic compliance, is performed to logically merge the visual classification probability distribution and the logical proof confidence result, thereby calculating and generating the final fusion detection result. Finally, the three-dimensional physical envelope coordinates, dynamic motion trends, and category attribute labels verified by physical axioms of the target object are characterized by the fusion of detection results.

Images