An automatic driving environment perception method

Abstract

This application provides an autonomous driving environment perception method that can accurately identify various types of obstacles in low-speed parking environments while maintaining the real-time performance and end-to-end inference advantages of a lightweight detection framework. It constructs a target focusing module, which uses a scale-aware dynamic strip convolution module to build serial horizontal and vertical depth convolutions based on adaptive kernel length, and incorporates directionally constrained offsets to fit corner and polygonal structures. Simultaneously, it adds directionally constrained local offsets, restricting the offset direction along the strip's extension direction, adapting to various slender targets while maintaining lightweight design. A dynamic dilation rate dilated convolution module improves the location reliability of low-lying obstacles. A lightweight channel attention interaction mechanism is introduced in the branch attention interaction fusion module to achieve bidirectional complementary enhancement of slender target features and low-lying target features. Finally, joint gating improves the model's classification accuracy for various types of obstacles.

Description

Technical Field

[0001] This invention relates to the field of intelligent assisted driving technology, specifically to an autonomous driving environment perception method. Background Technology

[0002] With the development of intelligent driving and automatic parking technologies, vehicles face higher demands for the accuracy and stability of their perception of the surrounding environment in scenarios such as low-speed parking, navigating narrow passages, reversing into parking spaces, parallel parking, and cruising in underground parking garages. Visual perception solutions, due to their lower cost, richer information content, and ease of identifying target categories and spatial semantics, have become an important component of automatic parking and low-speed assisted driving systems. Deep learning-based target detection models, such as the YOLO series, SSD, Faster R-CNN, and Transformer-type detection networks, have been widely applied to visual perception-based target detection tasks involving vehicles, pedestrians, and traffic facilities. These models can extract image features through convolutional neural networks or attention mechanisms and output target categories and location boxes, thus providing environmental information for vehicle planning and control. However, in practical applications, it has been found that existing general-purpose target detection models still have significant shortcomings in low-speed parking environments compared to highway scenarios.

[0003] First, parking environments are typically characterized by narrow spaces, close target distances, frequent occlusion, complex lighting, and strong interference from ground textures. Vehicles need to be able to identify low-lying obstacles promptly at low speeds. Some key targets in parking scenarios are characterized by their low height, small size, and proximity to the ground, such as low-lying obstacles, parking locks, the bottom of traffic cones, and the edges of curbs. These targets occupy a small percentage of pixels in the image and are often confused with ground shadows, brick seams, water stains, parking lines, or reflective areas, leading to missed detections or fluctuations in confidence levels in detection models. For parking safety, even small low-lying obstacles can cause damage to the vehicle's chassis, bumpers, or tires; therefore, their detection stability has high safety value.

[0004] Secondly, parking scenarios contain numerous elongated or strip-shaped targets, such as curbs, speed bumps, parking space line edges, wall base lines, and wheel chocks. These targets typically appear as long strips in images, narrow in width but long in length, exhibiting clear directional continuity. Traditional convolutional detection models, during multiple downsampling and feature fusion processes, easily weaken the fine-grained edge information of these targets, leading to broken target boundaries, detection box jitter, local missed detections, or missegmentation of continuous structures into multiple unstable regions. For automated parking systems, structures such as curbs, speed bumps, and parking space boundaries not only affect obstacle detection but also influence the determination of the vehicle's drivable area and parking trajectory planning.

[0005] Secondly, parking vision systems typically employ wide-angle or fisheye cameras to obtain a large near-field field of view. While these cameras can cover the area surrounding the vehicle, they are prone to radial distortion, perspective stretching, and boundary blurring at image edges and in close-range areas. Targets such as pillars, wall corners, close-range vehicle edges, and curb corners may deform in the image, leading to detection box positioning shifts, inaccurate edge recognition, or false detections of other categories. Especially when the vehicle is approaching obstacles at low speed, the target deformation changes rapidly with distance, further increasing the difficulty of stabilizing the detection model's output.

[0006] Furthermore, vehicle parking perception systems are typically deployed on embedded computing platforms, which are subject to strict limitations in terms of computing power, power consumption, latency, and cost. While large detection networks, complex attention modules, or multimodal fusion models can improve detection accuracy to some extent, their computational overhead is significant, making it difficult to meet the real-time and resource requirements of mass-produced vehicles. Summary of the Invention

[0007] To address the issue of insufficient accuracy in recognizing low-lying obstacles when applying existing vision-based target detection models to low-speed parking environments, this invention provides an autonomous driving environment perception method that can accurately identify various types of obstacles in low-speed parking environments while maintaining the real-time performance and end-to-end inference advantages of a lightweight detection framework.

[0008] The technical solution of the present invention is as follows: an autonomous driving environment perception method, characterized in that it includes the following steps: S1: Construct a scale-aware dynamic strip convolution module; The scale-aware dynamic bar convolution module includes: a lightweight feature morphological statistics layer and a dynamic one-dimensional decomposition convolution layer connected in sequence. The lightweight feature morphological statistics layer applies scale features to the input. Figure X s Perform global statistics and output the dynamic kernel length coefficient k. s And feed it into the dynamic one-dimensional decomposition convolution; k s =F stat (X s ); In the formula, F stat For dynamic kernel length coefficient k s A lightweight feature morphological statistics layer; s is the scale corresponding to the input image; The dynamic one-dimensional decomposition convolution is based on the dynamic kernel length coefficient k. s Construct sequential horizontal and vertical depthwise convolutions, and add orientation-constrained offsets to output the extracted bar features F. long,s Dynamic one-dimensional decomposition convolution includes the following operations: offset = F offset (X s ); F hori,s =DWConv 1×k_s,offset (X s ); F long,s =DWConv k_s×1,offset (F hori,s ); In the formula, DWConv is a depthwise separable convolution, 1×k s Represents adaptive horizontal bar convolution, k s ×1 represents adaptive vertical stripe convolution; offset is the orientation-constrained offset; F hori,s To calculate the intermediate value; F offset A lightweight feature mapping layer oriented towards orientation-constrained offsets; S2: Construct a dynamically dilated, dilated convolutional module; The dynamic dilation rate dilated convolution module is designed for input feature X. s Adaptive selection of expansion rate d s It dynamically matches the context range of obstacles of different sizes and outputs the context features F of low obstacles. small,s ; d s =F scale (X s ); F small,s =DWConv 3×3,d (X s ); In the formula, d takes the value d s ;F scale For the expansion rate d s A lightweight feature-scale statistical layer; S3: Construct a branch attention interaction fusion module; The branch attention interaction fusion module achieves bidirectional complementary enhancement of features of slender targets and features of low-lying targets through a lightweight channel attention interaction mechanism, and outputs the fused feature F. fuse,s ; Att long =CA(F long,s )⊙F small,s ; Att small =CA(F small,s )⊙F long,s ; F fuse,s = F long,s + Att long + Fsmall,s + Att small ; In the formula, CA(·) is the lightweight channel attention module, and ⊙ represents element-wise multiplication; Att long Features extracted for slender targets; Att small Features extracted for low-lying targets; S4: Constructing a joint gating system G s The joint gating G s Based on pre-generated dual prior parameters W dist and W pos Enhance distorted regions and suppress invalid backgrounds; G s =W dist ⊙W pos ; In the formula, W dist For the distortion weights acquired from the input image; W pos For perspective position weights acquired from the input image; S5: Construct the residual module; the residual module uses a residual structure to ensure gradient flow, and applies gated weights to the fused features to obtain the enhanced features F. out,s ; F out,s =X s +Gs⊙F fuse,s ; S5: Construct the target focusing module; The target focusing module includes: a scale-aware dynamic strip convolution module, a dynamic dilation rate dilated convolution module, a branch attention interaction fusion module, and a joint gating G... s and the residual module; Input feature X s The features are fed into the scale-aware dynamic strip convolution module and the dynamic dilation rate dilated convolution module, which are set in parallel. The features output by the two modules are simultaneously fed into the branch attention interaction fusion module for complementary fusion, and then sequentially fed into the joint gating G. s After processing in the residual module, the enhanced feature F is output. out,s ; S6: Construct an environment perception model based on a lightweight YOLO model; The environmental perception model includes: a backbone network, a neck network, a target focusing module, and a detection head; The backbone network, neck network, and detection head are implemented based on the standard structure of the YOLO model. A target focusing module is inserted between each detection head and the neck network. The target focusing module enhances the feature data output by the neck network, and then the enhanced feature F is processed.out,s It is then fed into the corresponding detection head for subsequent target detection.

[0009] Its further features are: In step S1, the dynamic kernel length coefficient k is... s Lightweight feature morphological statistical layer F stat and a lightweight feature mapping layer F with orientation-constrained offset offset The structures are the same, including: sequentially connected global average pooling and two fully connected layers; In step S2, the expansion rate d is... s Lightweight feature-scale statistical layer F scale It includes: a globally average pooling GAP layer connected in sequence, two fully connected layers connected in sequence, and three parallel classification layers; The input feature map is fed into a global average pooling (GAP) layer to extract features G representing the target scale and global density statistics of the entire feature map. Two fully connected layers are sequentially connected, and feature G is fed into the first fully connected layer FC1 to refine and compress the scale features F1. F1 = σ(FC1(G)); where σ is the ReLU activation function; The compressed scale feature F1 is then fed into the second fully connected layer FC2, which consists of three neurons and is mapped to three scores S corresponding to the three dilation rates. S = FC2(F1); Three classification layers are connected in parallel, and output scores S corresponding to low-dense targets, medium targets, and low-sparse targets, respectively. In the classification layer, based on a pre-set discrete inflation rate, the neuron index with the highest score is selected using argmax and mapped to the final inflation rate d. s :d s =argmax(S)+1; In step S4, the distortion weight W dist The calculation method is as follows: W dist =σ(λ1·r(x,y)); Perspective position weight W pos The calculation method is as follows: W pos =σ(λ2·y / H); In the formula, r(x,y) represents the normalized radial distance corresponding to pixel (x,y); r(x,y)=sqrt((xW / 2) 2 +(yH / 2) 2 ) / (sqrt(W 2 +H 2 ) / 2); sqrt() is the arithmetic square root function; σ is the sigmoid activation function; λ1 and λ2 are learnable hyperparameters; H is the scaling feature. Figure X s Image height, W is the scale feature. Figure X s The width; The environmental perception model also includes: a correction module; The correction module is located after the output of the detection head and before the non-maximum suppression (NMS) process. The correction module corrects the prediction result output by the detection head, performs subsequent NMS processing on the corrected prediction result, and then outputs the final result. The correction module includes: a near-range distortion sensing correction module, a small target soft distillation guidance module, a difficult target confidence compensation module, and a correction output module connected in sequence. The near-field distortion perception and correction module addresses the significant radial distortion produced by the parking wide-angle lens at close range by establishing a physical compensation model; it also establishes a near-field distortion perception and correction mapping function T(·) to correct the bounding box coordinate offset and outputs the corrected bounding box (bbox). i corr ; The difficult target confidence compensation module designs a confidence compensation mechanism to identify and correct such targets for low-confidence difficult samples, and outputs the compensated target confidence score (conf). i corr ; To address the issue of naturally low confidence levels for low-lying obstacles, the small-target soft distillation guidance module uses high-quality teacher signals to perform soft distillation constraint optimization on the target category, outputting a corrected category prediction vector cls. i corr ; The corrected output module constructs the corrected prediction result Output from the set of final perception results corresponding to all targets within a frame. crre The result is then fed into a subsequent module for non-maximum suppression (NMS) calculation. Output crre ={Out1,Out2,...,Out i ,…,Out M}; The corrected output for the i-th target is: Out i =(argmax(cls i corr ), bbox i corr ,conf i corr ); In the formula, M is the total number of targets detected in the input image of this frame; argmax(cls i corr This indicates that the index of the maximum value of the corrected category vector is taken as the final target category ID; In the near-field distortion perception and correction module, a near-field distortion perception and correction mapping function T(·) is constructed to correct the box coordinate offset for radial distortion of the parking near-field camera; bbox i corr =T(bbox i D i ); bbox i corr D represents the corrected target bounding box corresponding to the i-th target; i Let be the distance between the i-th target and the vehicle; The following operations are performed in the near-range distortion sensing correction mapping function T(·): a1: Determine D i Does it fall within the calibration range? If the error does not fall within the calibration range, no calibration operation will be performed. If it falls within the correction range, then perform the following steps; a2: Radial distance calculation: r = sqrt((cx - W / 2)) 2 +(cy-H / 2) 2 ) / (sqrt(W 2 +H 2 ) / 2); In the formula, r is the normalized radial distance; sqrt() is the arithmetic square root function; cx and cy are the pixel coordinates of the center point of the current target bounding box; W and H are the width and height of the input image; a3: Calculation of radial distortion stretching factor: radial_factor=1+k1*r 2 +k2*r 4 ; k1 and k2 are the camera radial distortion coefficients; radial_factor is the radial distortion stretching coefficient; a4: Bounding box width and height compensation: bbox_w corrected =bbox_w / radial_factor; bbox_h corrected =bbox_h / radial_factor; In the formula, bbox_w and bbox_h are the width and height of the original bounding box output by the model prediction head; bbox_w correctedTo compensate for the distortion of bbox_w, bbox_h corrected This is the height of the bounding box after distortion compensation for bbox_h; a5: Bounding box center point position compensation: c x_corrected =cx·(1+△ cx ·r); c y_corrected =cy·(1+△ cy ·r); In the formula, △ cx and △ cy It is the target bounding box center point offset compensation coefficient; c x_corrected and c y_corrected The coordinates of the center point of the frame after compensation; a6: Adjust the corrected bounding box width (bbox_w) corrected Corrected bounding box height bbox_h corrected and the coordinates of the center point of the bounding box (c x_corrected ,c y_corrected ) constitute bbox i corr The corresponding values; The small target soft distillation guidance module includes the following operations: b1: Construct a small target mask to filter the distillation target; small_mask=(bbox_area <Sdt); In the formula, bbox_area is the pixel area of ​​the bounding box of the current i-th target, bbox_area=w i ×h i Sdt is the threshold for small target determination; small_mask is a boolean mask for small targets; b2: Construct the KL divergence distillation loss formula; L ~ kd =KLDivLoss(F.log_softmax((cls_student) / T), F.softmax((cls_teacher) / T) )×T 2 ; In the formula, L ~ kd KLDivLoss is the KL divergence distillation loss; KLDivLoss is the function for calculating the KL divergence loss. cls_student represents the student model, and cls_teacher represents the soft label for the teacher model; F.log_softmax is the logarithmic softmax activation function, and F.softmax is the standard softmax activation function; T is the distillation temperature; T 2 This is a temperature coefficient compensation term, used to offset the effect of distillation temperature on the loss amplitude; b3: Mask Filtering: Distillation loss is filtered by a small target mask, retaining only the loss of small targets, and the average value of the retained small target losses is calculated to obtain L. kd ; L kd =(∑L ~ kd,i ·small_mask) / N; In the formula, N is the total number of samples in a single batch; L ~ kd,i This represents the divergence distillation loss corresponding to the i-th target within this batch; b4: Construct the total loss function L of the model total ; L total =L det +DLW×L kd ; In the formula, L det DLW represents the native detection loss output by the detection head of the environmental perception model; DLW represents the distillation loss weights. b5: Based on soft label fusion, the category prediction results are corrected; The training phase integrates teacher model knowledge to optimize category output: cls i corr = softmax( cls i + α·(cls teacher - cls i ) ); In the formula, α is the distillation-fusion weight; cls i corr This is the corrected category prediction vector, used for classification output; cls i This is the original prediction vector for the i-th category output by the detection head of the environmental perception model; The confidence compensation module for difficult targets includes the following operations: c1: Construct the aspect ratio prior factor ar_prior; ar_prior=σ(w i / (h i +10 -6 ) - Apt); In the formula, wi and h i The width and height of the original bounding box corresponding to the i-th target; 10 -6 To prevent the division into zero minimum values; σ is the Sigmoid activation function; Apt is the preset aspect ratio prior threshold; c2: Construct the location prior factor pos_prior; Calculate the normalized ordinate c' of the center point pixel of the current target bounding box. y :c' y =c y / H; where H is the height of the input image; Compare c'y with the preset position prior threshold Lpt; When c' y When Lpt is greater than 1, pos_prior = 1.0; When c' y When Lpt is less than or equal to Lpt, pos_prior = 0.5; c3: Construct the comprehensive compensation coefficient △conf; △conf i =△×ar_prior×pos_prior; In the formula, △ represents the preset maximum compensation range; △conf i This represents the comprehensive compensation coefficient corresponding to the i-th objective; c4: For the i-th detection target, determine whether the following four conditions are met simultaneously; if they are met simultaneously, execute c5 to perform confidence compensation; otherwise, do not perform confidence compensation for difficult targets, and directly proceed to the next step; Condition 1: The target category belongs to the hard target set HARD_CLASSES; The set of difficult targets, HARD_CLASSES, includes: low obstacles, curbs, and speed bumps; Condition 2: Initial confidence level (conf) i Condition met: Confidence lower limit threshold θ low <conf i <Confidence threshold θ high ; Condition 3: The target is elongated; the method for determining the elongated shape is that the aspect ratio of the target bounding box is greater than the prior threshold Apt for the aspect ratio. Condition 4: The target is located in the near-ground region of the image: the normalized ordinate c' of the center pixel of the current target bounding box. y Location prior threshold Lpt; c5: Compensated confidence level (conf) corr The calculation method is as follows: conf i corr=max (conf i +△conf i , θ low And conf i corr ≤0.9; In the formula, θ low The preset confidence level is the safety lower bound; conf i corr Let be the confidence level after final compensation for the i-th objective.

[0010] This application provides an autonomous driving environment perception method, which constructs a target focusing module. Within the target focusing module, a scale-aware dynamic strip convolution module is used, based on an adaptive kernel length k. s A serial horizontal and vertical depthwise convolution is constructed, while adding directionally constrained local offsets. The offset direction is constrained along the elongated extension direction to fit corner and polygonal structures, adapting to various shapes of slender targets while maintaining lightweight design. A dynamic dilation rate dilated convolution module is used to dynamically match the context range of obstacles of different sizes based on the pixel proportion of low-profile targets within the feature map, taking into account both the target body and contextual features such as ground shadows and water stains, thus improving the location reliability of low-profile obstacles. Then, a lightweight channel attention interaction mechanism is introduced into the branch attention interaction fusion module to achieve bidirectional complementary enhancement of slender target features and low-profile target features. Finally, a joint gating G... s This method adaptively enhances near-ground regions and areas with strong edge distortion, while adaptively suppressing invalid backgrounds such as the sky and garage roof, effectively increasing the model's classification accuracy for various types of obstacles. This approach can accurately identify various types of obstacles in low-speed parking environments while maintaining the real-time performance and end-to-end inference advantages of a lightweight detection framework. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall structure of the environmental perception model in this application; Figure 2 A schematic diagram of the target focusing module; Figure 3 This is a schematic diagram of the calibration module. Figure 4 For the expansion rate d s A schematic diagram of the structure of a lightweight feature-scale statistical layer; Figure 5 Ablation comparison for key sensing indicators; Figure 6 For comparison of the decrease in safety-related error indicators; Figure 7 To train simulation curves; Figure 8Comparison of AP@0.5 for six parking target categories; Figure 9 A comprehensive comparison of accuracy, speed, and parameter quantity; Figure 10 Normalized radar chart for parking perception capabilities. Detailed Implementation

[0012] This application includes an autonomous driving environment perception method, which includes the following steps.

[0013] To address the issues of poor generalization in traditional fixed-size one-dimensional convolution, the inability of manually set receptive fields to adapt to diverse and slender targets, and the lack of integration of fisheye imaging characteristics into simple region gating, such as... Figure 2 As shown, this application constructs a target focusing module, Parking-LoRA, for feature enhancement of low-profile and slender targets; specific details are shown below.

[0014] Most existing technologies rely on manually specified fixed-length one-dimensional convolutions, which can only capture standard straight-line slender structures and have poor extraction effects on corners and bends in kerbstones. This application abandons the method of manually fixing 1*7 and 7*1 convolution kernels and adopts dynamic convolution kernels.

[0015] S1: Construct a scale-aware dynamic strip convolution module. The scale-aware dynamic strip convolution is designed. The effective length of the convolution kernel is learned adaptively from the feature's own shape and scale. The direction-restricted local offset is added. It can adapt to various slender targets such as straight lines, L-shaped corners, and broken lines under the premise of lightweight design.

[0016] Define the input to the scale-aware dynamic bar convolution module: Let the batch size be B, the number of feature channels be C, and the height and width of the feature map be H and W, respectively. The module input is the original multi-scale feature X. s :X s The dimensions are B×C×H s ×W s ;s∈{3,4,5}; s=3: Corresponds to scale N3, step size 8, H3=W3=80; s=4: corresponds to scale N4, step size 16, H4=W4=40; s=5: corresponds to scale N5, step size 32, H5=W5=20.

[0017] The scale-aware dynamic bar convolution module includes: a lightweight feature morphostatistic layer and a dynamic one-dimensional decomposition convolution connected in sequence; Lightweight feature morphostatistical layer for input scale features Figure X s Perform global statistics and output the dynamic kernel length coefficient k. s And feed it into a dynamic one-dimensional decomposition convolution; ks =F stat (X s ); In the formula, F stat For dynamic kernel length coefficient k s Lightweight feature morphological statistical layer; k s k is obtained through network training and optimization. s ∈[3,7], where s is the scale corresponding to the input image; in this embodiment, s takes the value of 3, 4 or 5; Dynamic one-dimensional decomposition convolution is based on dynamic kernel length coefficient k s Construct sequential horizontal and vertical depthwise convolutions, add orientation-constrained offsets, fit corner and polygonal structures, and output the extracted bar feature F. long,s Dynamic one-dimensional decomposition convolution includes the following operations: offset = F offset (X s ); F hori,s =DWConv 1×k_s,offset (X s ); F long,s =DWConv k_s×1,offset (F hori,s ); In the formula, DWConv is a depthwise separable convolution, 1×k s This indicates an adaptive horizontal strip convolution that captures the edges of lateral curbs and speed bumps; k s ×1 indicates adaptive vertical strip convolution, enhancing the continuity of the vertical structure; offset is a direction-constrained offset, only slightly shifting along the extension direction of the strip, specifically adapted for L-shaped and polygonal curb corners; F hori,s To calculate the intermediate value; F offset It is a lightweight feature mapping layer oriented towards directionally constrained offsets.

[0018] Specifically, this applies to dynamic kernel length coefficient k. s Lightweight feature morphological statistical layer F stat and a lightweight feature mapping layer F with orientation-constrained offset offset The structure is the same, including: a globally average pooling layer and two fully connected layers connected in sequence; the scale-aware dynamic bar convolution designed in this application has an effective length of the dynamic convolution kernel that is autonomously learned by the feature statistics layer.

[0019] S2: Construct a dynamically dilated convolutional module. Using dynamically dilated convolution, the receptive field is adaptively adjusted based on the pixel proportion and distribution of low-lying targets within the feature map, and the dilation rate d is adaptively selected. s∈{1,2,3}, and then dynamically match the context features of obstacles of different sizes.

[0020] The dynamic dilation rate dilated convolution module targets the input feature X. s Adaptive selection of expansion rate d s It dynamically matches the context range of obstacles of different sizes and outputs the context features F of low obstacles. small,s ; d s =F scale (X s ); F small,s =DWConv 3×3,d (X s ); In the formula, d takes the value d s ;F scale For the expansion rate d s A lightweight feature scale statistical layer.

[0021] like Figure 4 As shown, facing the expansion rate d s Lightweight feature-scale statistical layer F scale It includes: a globally average pooling GAP layer connected sequentially, two fully connected layers, and three classification layers; wherein, the two fully connected layers are connected sequentially, and the three classification layers are connected in parallel, corresponding to the output d of low-density targets, medium-density targets, and low-density sparse targets, respectively. s ; Ensure that the dilation rate d can be adaptively selected based on the pixel proportion of low-lying targets within the feature map. s .

[0022] Specifically, by using Global Average Pooling (GAP), the global spatial dimension of the entire image features is compressed, and features G that represent the target scale and global density statistical information of the entire feature map are extracted: G = GAP(Xs) ∈ R. B×C Operation: For H s W s The average value of the dimensions is calculated, spatial location is removed, and the global semantics of the channel dimension are preserved; the overall size ratio and density of low-rise targets in the current area are statistically analyzed.

[0023] Then, dimensionality reduction and nonlinear activation are performed through the first fully connected layer, and feature compression is performed using a very small number of neurons to further refine scale features: F1=σ(FC1(G)); FC1 is the first fully connected layer, which significantly reduces the dimensionality of the channels, such as reducing the number of channels C to 16, and fixing the number of channels as a hyperparameter to achieve lightweight design. σ is the ReLU activation function, which introduces nonlinearity.

[0024] The second fully connected layer maps the features to classification scores S. The compressed features are then mapped to three-class scores, corresponding to inflation rates of 1, 2, and 3.

[0025] S=FC2(F1)∈R B×3 ; In the formula, FC2 is the second fully connected layer, the number of neurons Neu is fixed at 3, and the index of the neurons Neu is {0,1,2}; the output of Neu0 corresponds to the expansion rate d. s =1 score S0; Neu1 output corresponds to expansion rate d s =2 score S 1、 Neu2 outputs the corresponding expansion rate d s =3 for the score S2; S=[ S0, S1, S2]; No activation function, directly output the original classification logits.

[0026] Finally, in the classification layer, sampling is performed based on a preset discrete inflation rate. The index with the highest score is selected using argmax, and the category with the highest score is mapped to the final inflation rate d. s ∈{1,2,3}, and are fed into subsequent dynamic dilated convolution; d s =argmax(S)+1; Assume that S = FC2(F1) = [S0, S1, S2] = [0.2 points, 0.7 points, 0.1 points]; d s =argmax(S)+1=1+1=2, meaning the neuron with index 1 outputs the highest score, and the dilation rate d of the input feature map is... s The value is 2.

[0027] Based on the proposed solution, the expansion rate is automatically reduced when the target is dense and small in size, and automatically increased when the target is sparse and large in size, taking into account the contextual features of the target body and ground shadows, water stains, etc., to improve the reliability of positioning low obstacles.

[0028] Because the two modules collect different types of features, directly adding the outputs of the dynamic dilation rate dilated convolution module and the scale-aware dynamic strip convolution module would result in feature independence and information fragmentation. Therefore, this application does not use the direct addition method.

[0029] S3: Construct a branch attention interaction fusion module, introduce a branch attention interaction fusion mechanism to achieve bidirectional information complementarity between long and narrow target branches and small target branches, and solve the problem of feature fragmentation between two branches.

[0030] The branch attention interaction fusion module uses a lightweight channel attention interaction mechanism to achieve bidirectional complementary enhancement of features of slender targets and features of low-lying targets, and outputs the fused feature F.fuse,s ; Att long =CA(F long,s )⊙F small,s ; Att small =CA(F small,s )⊙F long,s ; F fuse,s = F long,s + Att long + F small,s + Att small ; In the formula, CA(·) represents the lightweight channel attention module, and ⊙ represents element-wise multiplication. Att long Features extracted for slender targets; Att small Features extracted for low-lying targets; The two branches extract each other's effective features, and after fusion, they retain both the slender structure and the contextual information of the small target.

[0031] S4: Constructing a joint gating system G s This method combines the physical characteristics of radial distortion from a parking fisheye camera with prior knowledge of ground perspective position to construct a dynamic spatial weighted graph G. s The weights change in real time with pixel position and distortion intensity. By using a dual-prior joint gating of distortion and position, the radial distortion intensity of the fisheye and the ground perspective position information are fused to replace the traditional fixed-area hard gating, adaptively enhancing high-risk areas such as near the ground and edges of the image.

[0032] Joint Gating G s Based on pre-generated dual prior parameters W dist and W pos Enhance distorted regions and suppress invalid backgrounds; G s =W dist ⊙W pos ; In the formula, W dist For the distortion weights acquired from the input image; W pos G represents the perspective position weights acquired from the input image. s The dimensions are B×1×H s ×W s B represents the batch. Passing through G s It adaptively enhances the near-ground region and areas with strong edge distortion of the image, and adaptively suppresses invalid backgrounds such as the sky and the top of the garage.

[0033] In step S4, the distortion weight W dist The calculation method is as follows: Wdist =σ(λ1·r(x,y)); W dist The strength of fisheye distortion is represented by the weight of the image edge distortion, which is higher when the edge distortion is stronger. Perspective position weight W pos The calculation method is as follows: W pos =σ(λ2·y / H); W pos To determine perspective position, the closer a pixel is to the ground (the larger its y-value), the higher its weight.

[0034] In the formula, r(x,y) is the normalized radial distance corresponding to pixel (x,y), which represents the distortion intensity; r(x,y)=sqrt((xW / 2) 2 +(yH / 2) 2 ) / (sqrt(W 2 +H 2 ) / 2); sqrt() is the arithmetic square root function; σ is the Sigmoid activation function, with an output range of [0,1]; λ1 and λ2 are learnable hyperparameters; H is the scaling feature. Figure X s Image height, W is the scale feature. Figure X s The width.

[0035] S5: Constructing the residual module; the residual module uses a residual structure to ensure gradient flow, and applies gated weights to the fused features to obtain the enhanced features F. out,s ; F out,s =X s +Gs⊙F fuse,s .

[0036] The enhanced feature X', output from all three scales, is fed into the detection head. X'3=F out,3 ; X'4=F out,4 ; X'5=F out,5 .

[0037] S5: Construct the target focusing module (Parking-LoRA); like Figure 2 As shown, the target focusing module includes: a scale-aware dynamic strip convolution module, a dynamic dilation rate dilated convolution module, a branch attention interaction fusion module, and a joint gating G module. s and residual modules; Input feature X sThe features are fed into a scale-aware dynamic strip convolution module and a dynamic dilation rate dilated convolution module, which are set in parallel. The features output by the two modules are simultaneously fed into a branch attention interaction fusion module for complementary fusion, and then sequentially fed into a joint gating G. s After processing in the residual module, the enhanced feature F is output. out,s .

[0038] Parking-LoRA no longer relies on manually set fixed one-dimensional convolutional kernels. It adaptively matches slender targets of different scales and shapes through scale-aware dynamic strip convolutions, effectively solving the problem of feature fragmentation in zigzag curbs and speed bumps. Dynamic dilated convolutions can flexibly adapt to the contextual requirements of various low-lying obstacles. Branch attention interaction fusion allows two types of features to mutually enhance each other. Combined with dual prior gating of fisheye distortion and perspective position, it conforms to the imaging rules of parking cameras, accurately focusing on high-risk areas near the ground. Multiple mechanisms work synergistically to further enhance the feature representation capabilities of low-lying obstacles and slender strip targets while ensuring the model's lightweight and real-time performance. The entire target focusing module of this method uses depthwise separable convolutions, with the increase in the number of new parameters and inference latency both controlled within 2%, maintaining the model's lightweight and easy-to-deploy characteristics.

[0039] The target focusing module Parking-LoRA constructed in this application is a plug-and-play neck feature enhancement module. Inserted after the neck output and before the detection head, it specifically enhances the features of three core difficult targets: low obstacles, curbs, and speed bumps. The number of parameters increases by ≤ 2%, and the inference FPS loss is ≤ 2%. The target focusing module constructed in this application does not add any additional deployment burden to the lightweight architecture and has almost no impact on its original high-efficiency inference characteristics.

[0040] S6: Construct an environment perception model based on a lightweight YOLO model; the environment perception model includes: backbone network, neck network, target focusing module Parking-LoRA, detection head and correction module Parking-Distill.

[0041] The target focusing module constructed in this application is applicable to YOLO series models. In this embodiment, an environment perception model is constructed based on the YOLOv10S model.

[0042] Specifically, such as Figure 1 As shown, the backbone network, neck network, and detection head are implemented based on the standard structure of the YOLO model, with a target focusing module inserted between each detection head and the neck network. The target focusing module enhances the feature data output by the neck network, and then the enhanced features F are processed. out,s It is then fed into the corresponding detection head for subsequent target detection.

[0043] The YOLOv10s backbone uses a lightweight CSPDarkNet variant, combined with the C2f (Cross Stage Partial Faster) module, to downsample the 640×640 input image layer by layer, outputting three scale feature maps for subsequent processing. The backbone network analyzes key feature layers for parking scenarios: The scale feature map P3 has a resolution of 80×80, a stride of 8, a receptive field of ~52×52px, and the main targets detected are: low obstacles and local curb stones. The scale feature map P4 has a resolution of 40×40, a stride of 16, and a receptive field of ~104×104px. The main targets detected are speed bumps and garage pillars. The scale feature map P5 has a resolution of 20×20, a stride of 32, and a receptive field of ~208×208px. The main targets detected are vehicles, walls, and parking signs. Low-lying obstacles (height <30cm) are primarily detected using the P3 feature layer. When the target height accounts for only 3% to 8% of the image height, the 80×80 feature map of the P3 layer provides the finest spatial resolution and is the layer that the subsequent Parking-LoRA module focuses on enhancing.

[0044] The YOLOv10s Neck network employs a bidirectional feature pyramid structure, which performs bidirectional fusion of the backbone outputs P3 / P4 / P5 in a top-down (semantic enhancement) and bottom-up (detail enhancement) manner, outputting three sets of fused feature maps: N3, N4, and N5.

[0045] N3 (80×80): P3 fine-grained texture + P4 mid-layer semantics, mainly for low-lying obstacles; N4 (40×40): P4 mid-level features + P5 high-level semantics, for speed bump and pillar detection; N5 (20×20): High-level semantic features, targeting large targets such as vehicles and buildings.

[0046] After Neck outputs N3, N4, and N5, they are sent to the Parking-LoRA module for targeted enhancement.

[0047] The YOLOv10s model uses decoupled detection heads, separating classification and bounding box regression into two independent branches to avoid gradient interference and improve the accuracy of small target detection. The specific details of the three detection heads are as follows: The detection head has an N3 detection scale, a grid number of 80×80=6400, a step size of 8px, and predicts target size <80×80px; it is mainly used for curbstone cut surfaces and small obstacles. The detection head has an N4 detection scale, a grid number of 40×40=1600, a step size of 16px, and predicts target size of 80~160px; it mainly targets targets such as speed bump width and column cross-section. The detection head has an N5 detection scale, a grid number of 20×20=400, a step size of >160px, and a predicted target size of <80×80px; it is mainly used for targets such as vehicle outlines, walls, and door frames.

[0048] Parking scenarios typically utilize wide-angle or fisheye cameras, covering the immediate surroundings of the vehicle. In this embodiment, the captured image quality details are as follows: Image resolution: 1280×720 / 1920×1080, balancing clarity and transmission bandwidth; Field of view (FOV): ≥120° horizontally, covering near blind spots; Frame rate ≥ 30FPS, meeting real-time perception requirements; Lens type: Wide-angle / fisheye (f<2.0mm) will cause near-field distortion; Installation location: below the front bumper / rearview mirror, capturing images from a low angle, close to the ground target; Dynamic range ≥100dBHDR, suitable for strong light and low light conditions in underground parking garages.

[0049] In practice, the acquired images need to be preprocessed. Preprocessing includes: scaling the input image (any resolution) acquired from the vehicle-mounted camera to 640×640 proportionally, maintaining the aspect ratio, and filling any insufficient areas with gray (pixel value 114) to avoid distortion of the proportions of slender objects (curb stones). After converting from BGR to RGB, the pixel values ​​are normalized to [0.0, 1.0], and then standardized according to ImageNet mean and variance.

[0050] During the training phase, data augmentation is performed. The augmentation types and parameters are shown in Table 1. Table 1: Examples of Data Augmentation Rules During Training

[0051] This application also includes a correction module, Parking-Distill, designed for low-speed parking scenarios (vehicle speed < 30 km / h) in autonomous driving. This module is deployed after the detection head output and before NMS (Non-Maximum Suppression). It corrects the prediction results output by the detection head, performs NMS processing on the corrected prediction results, and then outputs the final result. The core features of the correction module are: effective constraints during the training phase, zero computational overhead during the inference phase, adaptability to the computing power and latency limitations of automotive embedded platforms, and specifically addresses issues such as near-range radial distortion from fisheye / wide-angle cameras, low confidence levels for low / slender targets, and false positives and false negatives for small targets.

[0052] The input to the Parking-Distill calibration module is the raw prediction results (class, bounding box, confidence score) output from the YOLOv10s decoupled detection head; for example... Figure 3 As shown, the correction module includes: a near-range distortion perception correction module, a small-target soft distillation guidance module, a difficult-target confidence compensation module, and a correction output module connected in sequence. These modules perform three optimization steps sequentially: near-range distortion perception correction, small-target soft distillation guidance, and difficult-target confidence compensation. The correction output module then outputs the corrected category, bounding box, and confidence score, which are then fed into NMS for final processing. The distillation and loss constraints in the correction module only exist during the model training phase; there is no additional computation or latency increase during model deployment and inference.

[0053] The near-field distortion perception and correction module addresses the significant radial distortion produced by the parking wide-angle lens at close range by establishing a physical compensation model; it also establishes a near-field distortion perception and correction mapping function T(·) to correct the bounding box coordinate offset and outputs the corrected bounding box (bbox). i corr ; The difficult target confidence compensation module designs a confidence compensation mechanism to identify and correct such targets for low-confidence difficult samples, and outputs the compensated target confidence score (conf). i corr ; The small-target soft distillation guidance module addresses the issue of naturally low confidence levels for low-lying obstacles by using high-quality teacher signals to perform soft distillation constraint optimization on the target category, outputting a corrected category prediction vector cls. i corr ; The corrected output module constructs the corrected prediction result Output from the set of final perception results corresponding to all targets within a frame. crre The result is then fed into a subsequent module for non-maximum suppression (NMS) calculation. Output crre ={Out1,Out2,...,Out i ,…,Out M}; The corrected output for the i-th target is: Out i =(argmax(cls i corr ), bbox i corr ,conf i corr ); In the formula, M is the total number of targets detected in the input image of this frame; argmax(cls i corrThe expression indicates that the index of the maximum value of the corrected category vector is used as the final target category ID.

[0054] Assume that the total number of detected targets in a single frame image output by the detection head is M, and the total number of target categories is K. For the i-th detected target in a single frame, the original input variables for the correction module are defined as follows: The detection head outputs the raw class prediction vector: cls i ∈R K ;cls i These are category Logits that have not undergone Softmax. The detector outputs the original bounding box: bbox i =[x i ,y i ,w i ,h i ]; The raw target confidence score output by the detection head is: conf i ∈R; Where i represents the i-th target, cls i Logits are categories that have not undergone Softmax.

[0055] W and H are the total width and total height of the input image.

[0056] Standard detection frame formats include: (x i ,y i ) represents the coordinates of the center point of the i-th target bounding box; w i ,h i The bounding box width and height are the target bounding box values; the coordinates can be based on image pixel normalization or original pixels.

[0057] In this embodiment, K is set to 6, including: low_obstacle, person, vehicle, curb, speed bump, and pillar.

[0058] In parking scenarios, standard wide-angle / fisheye cameras with a focal length f<2.0mm exhibit severe radial distortion at close ranges of 0.5~5m, leading to center point shifts and stretching deformation of the target bounding box. This method constructs a near-range distortion perception and correction module, performing geometric compensation based on the camera's optical distortion model to correct the box positioning error caused by distortion. All distortion parameters are obtained from offline camera calibration, and only linear calculations are performed during inference.

[0059] In the near-field distortion perception and correction module, a near-field distortion perception and correction mapping function T(·) is constructed to correct the box coordinate offset for radial distortion caused by the parking near-field camera. bbox icorr =T(bbox i D i ); bbox i corr D represents the corrected target bounding box corresponding to the i-th target; i The distance between the i-th target and the vehicle is obtained by the distance sensor in the vehicle parking perception system, such as the measurement data of the on-board ranging module, or calculated based on camera calibration information.

[0060] Perform the following operations in the near-field distortion sensing correction mapping function T(·): a1: Determine D i Whether it falls within the correction range; in this embodiment, the correction range is set to [0.5 m, 5 m]; If the error does not fall within the calibration range, no calibration operation will be performed. If it falls within the correction range, then perform the following steps a2~a6; a2: Radial distance calculation: r = sqrt((cx - W / 2)) 2 +(cy-H / 2) 2 ) / (sqrt(W 2 +H 2 ) / 2); In the formula, r is the normalized radial distance with a value range of [0,1]; the larger r is, the closer the pixel is to the image edge and the stronger the fisheye distortion.

[0061] sqrt() is the function to calculate the arithmetic square root; cx and cy are the pixel coordinates of the center point of the current target bounding box, c x =x i c y =y i W and H are the width and height of the input image; in this embodiment, W=H=640 pixels; numerator sqrt((cx-W / 2) 2 +(cy-H / 2) 2 ) represents the Euclidean distance from the center of the bounding box to the geometric center of the image; the denominator sqrt(W 2 +H 2 ) / 2 represents half the length of the image diagonal, and is the normalization coefficient.

[0062] a3: Calculation of radial distortion stretching factor: radial_factor=1+k1*r 2 +k2*r 4 ; k1 and k2 are the camera radial distortion coefficients, calculated offline using camera calibration tools. They are fixed constants during the inference phase and do not require dynamic updates. radial_factor is the radial distortion stretching coefficient, which is the significant radial distortion step size value generated by the parking wide-angle lens at close range. The value range of radial_factor is usually [0.8, 1.3]. The value is larger in the edge region, indicating more obvious stretching.

[0063] a4: Compensates for bounding box width and height, corrects stretching deformation: bbox_w corrected =bbox_w / radial_factor; bbox_h corrected =bbox_h / radial_factor; In the formula, bbox_w and bbox_h are the width and height of the original bounding box output by the model prediction head, that is, the value is taken as w output by the detection head. i and h i bbox_w corrected To compensate for the distortion of bbox_w, bbox_h corrected This is the height of the bounding box after distortion compensation for bbox_h.

[0064] a5: Compensates for the center point position of the bounding box, corrects center offset. c x_corrected =cx·(1+△ cx ·r); c y_corrected =cy·(1+△ cy ·r); In the formula, △ cx and △ cy This is the target bounding box center point offset compensation coefficient, obtained offline from camera calibration data, and is a globally fixed constant; c x_corrected and c y_corrected These are the coordinates of the center point of the frame after compensation.

[0065] a6: Adjust the corrected bounding box width (bbox_w) corrected Corrected bounding box height bbox_h corrected and the coordinates of the center point of the bounding box (c x_corrected ,c y_corrected ) constitute bbox i corr The corresponding values.

[0066] The near-range distortion sensing and correction module only performs distortion correction on targets within 0.5~5m of the camera; for distant targets, the original bounding box is directly reused; all parameters k1, k2, Δ cx , △ cyAll values ​​are calibrated offline and are fixed, with no dynamic computational overhead for inference.

[0067] In parking scenarios, small targets such as low obstacles and traffic cones (box area < 32×32 pixels) have weak features and chaotic distribution of category prediction confidence. This module employs knowledge distillation, using a high-precision large model as the "teacher model" to guide the lightweight YOLOv10s (student model) to learn the category distribution of small targets. The distillation loss is calculated only during the training phase and does not take effect during the inference phase.

[0068] In this embodiment, the hyperparameters for the global configuration of the basic distillation parameters are as follows: Distillation temperature T=4 is used to soften the Softmax output, smooth the class probability distribution, and avoid overfitting to hard labels. Distillation loss weight DLW=0.5, used as a fixed coefficient to balance the proportion of detection loss and distillation loss; The small target determination threshold Sdt = 32 × 32 px, where Sdt is the area threshold. Only targets with a frame area smaller than this value participate in the distillation constraint.

[0069] The small-target soft distillation guide module includes the following operations: b1: Construct a small target mask to filter the distillation target; small_mask=(bbox_area <Sdt); In the formula, bbox_area is the pixel area of ​​the bounding box of the current i-th target, bbox_area=w i ×h i Sdt is the threshold for determining small targets; `small_mask` is a boolean mask for small targets. 1 represents True, indicating a small target that participates in distillation; 0 represents False, indicating a large target that does not participate in distillation.

[0070] b2: Construct the KL divergence distillation loss formula; L ~ kd =KLDivLoss(F.log_softmax((cls_student) / T),F.softmax((cls_teacher) / T))×T 2 ; In the formula, L ~ kd KLDivLoss is the KL divergence distillation loss; KLDivLoss is the KL divergence loss calculation function, which measures the difference between the distribution of student and teacher categories. cls_student is the student model, which in this example is YOLOv10s, with the original category Logits, i.e., the input variable cls.i ; cls_teacher is the soft label for the teacher model. In this embodiment, the teacher model uses the high-resolution inference results of YOLOv10m to ensure the quality of the soft label. F.log_softmax is the logarithmic softmax activation function, and F.softmax is the standard softmax activation function. These two functions are used for probability normalization. T is the distillation temperature; T 2 This is a temperature coefficient compensation term, used to offset the effect of distillation temperature on the loss amplitude.

[0071] b3: Masked Filtering: Distillation loss is filtered using a small-target mask, retaining only the loss of small targets, and the average value of the retained small-target losses is calculated to obtain L. kd ; L kd =(∑L ~ kd,i ·small_mask) / N; In the formula, N is the total number of samples in a single batch; L ~ kd,i L represents the divergence distillation loss corresponding to the i-th target in this batch; kd This represents the average loss of small targets within this batch. b4: Construct the total loss function L of the model total ; L total =L det +DLW×L kd ; In the formula, L det The native detection loss output by the detection head of the environment perception model includes: classification loss + bounding box regression loss + confidence loss; DLW is the distillation loss weight, which is set to 0.5 in this embodiment to balance the gradient contributions of the two losses.

[0072] b5: Based on soft label fusion, the category prediction results are corrected; During the training phase, teacher model knowledge is integrated to optimize category output. cls i corr = softmax( cls i + α·(cls teacher - cls i ) ); In the formula, α is the distillation-fusion weight, and the value range is α∈(0,1). In this embodiment, α=0.3. cls i corr This is the corrected category prediction vector, used for classification output; cls i This is the original prediction vector for the i-th category output by the detection head of the environment perception model.

[0073] The small-target soft distillation guidance module can transfer the fine-grained recognition capabilities of large models for difficult targets in parking lots to small vehicle-side models in a lightweight and targeted manner. This allows student models to better detect small, low, fine, and weak targets such as curbs, speed bumps, and low obstacles without increasing inference costs, while improving recall, reducing false positives, and meeting real-time deployment requirements. The small-target soft distillation guidance module is only effective during the training phase; distillation loss and class fusion calculations are not required during the inference phase. The small-target soft distillation guidance module is strictly limited to small targets with a bounding box area < 32×32 pixels, and does not interfere with large targets such as vehicles and pedestrians. During the training phase, the small-target soft distillation guidance module can make the class output more accurate.

[0074] Curbstones, speed bumps, and low obstacles are considered difficult targets with weak textures, resulting in generally low model output confidence scores (0.35~0.55), making them highly susceptible to being missed by NMS threshold filtering. This application constructs a difficult target confidence compensation module, which adaptively enhances the confidence of targets of specified difficulty categories, specific shapes, and locations. The parameters are globally fixed, and the computation is lightweight during the inference phase.

[0075] In this embodiment, the globally configured hyperparameters in the difficult target confidence compensation module are as follows: 1) Confidence threshold θ low =0.35, compensation is initiated only for targets with a confidence level greater than this value; 2) Confidence threshold θ high =0.55, compensation is initiated only for targets with a confidence level less than this value; 3) The maximum compensation range Δ=0.12, which is the maximum increase in confidence level in a single instance, to prevent the confidence level from being artificially inflated; 4) Aspect ratio prior threshold Apt=3.0, dimensionless, used for determining slender targets: if the box width / box height is greater than 3, it is determined to be a curb / speed bump; 5) Location prior threshold Lpt=0.65, normalized coordinates, used for near-ground region determination: if the center of the bounding box c_y>0.65, it means it is located in the lower 1 / 3 region of the image; 6) The set of difficult targets, HARD_CLASSES, is set to {low obstacles, curbs, speed bumps}. Confidence compensation is applied only to target types included in this set. The specific target types included in the difficult target set are predefined based on the actual vehicle operating scenario.

[0076] The confidence compensation module for difficult targets includes the following operations: c1: Construct the aspect ratio prior factor ar_prior to adapt to slender targets; ar_prior=σ(w i / (h i +10 -6 ) - Apt); In the formula, w i and h i The width and height of the original bounding box corresponding to the i-th target; 10 -6 To prevent the elimination of zero minimum values; σ is the Sigmoid activation function, with an output range of [0,1]. Apt is the aspect ratio prior threshold.

[0077] The thinner and longer the bounding box (aspect ratio > 3), the closer ar_prior is to 1, and the higher the compensation weight.

[0078] c2: Construct a location prior factor pos_prior to adapt to low-lying targets near the ground; Calculate the normalized ordinate c' of the center point pixel of the current target bounding box. y :c' y =c y / H; where H is the height of the input image; c' y Range [0,1]; Compare c'y with the preset position prior threshold Lpt; When c' y When Lpt is greater than 1.0, pos_prior = 1.0, indicating a high-risk area near the ground, and full weight enhancement is applied. When c' y When Lpt is less than or equal to Lpt, pos_prior = 0.5, in the upper background region, half-weighted weakening is applied.

[0079] The position weight factor pos_priord has only two fixed values: 0.5 and 1.0, representing the confidence compensation weight coefficient for the current target's location. When the target is located in the lower 1 / 3 of the image near the ground (a region with a high incidence of low obstacles), the compensation weight is maximized to 1.0; when the target is located in the upper region of the image, the weight is halved.

[0080] c3: Calculate the comprehensive compensation coefficient △conf; △conf i =△×ar_prior×pos_prior; In the formula, △ represents the preset maximum compensation range, which is 0.12 in this embodiment; △conf i This represents the comprehensive compensation coefficient corresponding to the i-th objective.

[0081] c4: For the i-th detection target, determine whether the following four conditions are met simultaneously; if they are met simultaneously, execute c5 to perform confidence compensation; otherwise, do not perform confidence compensation for difficult targets, and directly proceed to the next step; Condition 1: The target category belongs to the hard target set HARD_CLASSES; The set of difficult targets (HARD_CLASSES) includes: low obstacles, curbs, and speed bumps. Condition 2: Initial confidence level (conf) i Condition met: Confidence lower limit threshold θ low <conf i <Confidence threshold θ high ; Condition 3: The target has a slender shape; the method for determining a slender shape is that the aspect ratio of the target bounding box is greater than Apt. Condition 4: The target is located in the near-ground region of the image: the normalized ordinate c' of the center pixel of the current target bounding box. y Location prior threshold Lpt.

[0082] c5: Compensated confidence level (conf) corr The calculation method is as follows: conf i corr =max (conf i +△conf i , θ low And conf i corr ≤0.9; In the formula, θ low The preset confidence level safety lower bound is set to 0.35 in this embodiment; conf i corr Let conf be the final compensated confidence level corresponding to the i-th target; i corr The upper limit is forcibly set to 0.9 when conf i corr If the calculated value is greater than 0.9, it will be forcibly assigned the value of 0.9 to avoid false high confidence.

[0083] To verify the performance of this method, a comparative experiment was conducted with existing models. The comparison models included: YOLOv8s: a commonly used lightweight baseline; YOLOv10s: the base model of this method; YOLOv10s + Parking-LoRA target focusing module: used to verify the neck feature enhancement effect; YOLOv10s + Parking-LoRA target focusing module + Parking-Distill correction module: the complete model constructed in this application.

[0084] Dataset simulation settings include: Scenarios: underground parking garage, surface parking lot, narrow passage, reversing into a parking space, parallel parking, and slow detour.

[0085] Categories: low_obstacle, person, vehicle, curb, speed_bump, pillar.

[0086] Speed ​​range: 0-30 km / h.

[0087] Camera: Wide-angle / fisheye, horizontal FOV ≥ 120°, input resolution 640×640.

[0088] Disturbances: low light, reflections, shadows, water stains, near-range radial distortion, partial occlusion.

[0089] Evaluation indicators include: Precision categories: mAP50, Low Obstacle Recall, Curb AP@0.5, Speed ​​Bump AP@0.5, Curb F1-Score.

[0090] Safety-related: low obstacle miss rate, pillar / vehicle corner false detection rate, and close-range (<5m) positioning error.

[0091] Deployment-related factors: number of model parameters, FPS, and additional inference overhead.

[0092] For the accuracy-related experimental results, please refer to the appendix. Figure 5 The figure shows a performance comparison of YOLOv8s, YOLOv10s, YOLOv10s+Parking-LoRA, and the complete algorithm YOLOv10s+Parking-LoRA+Parking-Distill in detecting key targets during low-speed parking. The results show that the basic YOLOv10s already possesses better lightweight detection capabilities compared to YOLOv8s. Furthermore, the addition of Parking-LoRA significantly improves the recall for low obstacles, the AP@0.5 for curbs, and the AP@0.5 for speed bumps. Specifically, the recall for low obstacles increased from 61.2% to 75.8%, the AP@0.5 for curbs from 54.3% to 69.7%, and the AP@0.5 for speed bumps from 58.9% to 74.2%.

[0093] The results demonstrate that Parking-LoRA, through 1×7+7×1 asymmetric convolution, dilated convolution, and low-region gated attention, effectively enhances the feature representation of low-lying, small-sized, and elongated targets in the near-field region below the image. Further incorporating Parking-Distill improves the algorithm across four key metrics, particularly achieving an F1-Score of 84.1% for curb stones, indicating that the proposed algorithm not only improves recall but also enhances the stability and consistency of detection results. Figure 5 Data shows that the algorithm in this application enhances the features of low and slender targets through Parking-LoRA and improves low-confidence and distorted targets through Parking-Distill. This method has a significant enhancement effect on low obstacles, curbs and speed bumps that are most easily missed in parking scenarios, and can effectively alleviate the feature loss problem of general detection networks on low and slender targets.

[0094] Figure 6 From a safety perspective, the performance of different models was compared in terms of near-range localization error, false negative rate for low-lying obstacles, and false positive rate for pillars / vehicle corners. The results show that after adding Parking-LoRA, the false negative rate for low-lying obstacles decreased from 29.8% in the base model to 24.8%, indicating that the neck feature enhancement module can significantly improve the recall capability for small, low-lying targets. Further introducing Parking-Distill reduced the near-range localization error from 8.3px to 4.7px, a 43.4% decrease relative to the Parking-LoRA baseline; the false negative rate for low-lying obstacles decreased from 24.8% to 17.2%, a 30.6% decrease; and the false positive rate for pillars / vehicle corners decreased from 12.4% to 9.1%, a 26.6% decrease.

[0095] The results show that Parking-Distill significantly suppresses near-range radial distortion, edge stretching, and boundary blurring caused by wide-angle or fisheye parking cameras. Through near-range distortion perceptual correction, small-target soft distillation constraints, and difficult-target confidence compensation, the model can output more stable bounding boxes and confidence scores in complex near-range scenes. Figure 6 The comparison results show that this algorithm not only improves detection accuracy, but also directly reduces safety risk indicators such as missed detections, false detections, and positioning deviations, making it suitable for low-speed parking, close-range obstacle avoidance, and automatic parking assistance scenarios.

[0096] Figure 7The data shows the convergence curves of the validation set mAP50 and the trend of the total training loss for different models during the training process. The mAP50 curves show that the model with Parking-LoRA improves faster in the early and mid-stages of training, ultimately achieving a significantly higher convergence accuracy than the basic YOLOv10s. With the addition of Parking-Distill to the complete algorithm, the mAP50 curve shifts further upward, eventually reaching approximately 79.1%, indicating that soft distillation and distortion correction constraints provide more stable training supervision for the model.

[0097] The training loss curves show that the full algorithm exhibits a faster loss reduction rate and a lower final loss. This indicates that Parking-LoRA provides a more effective representation of parking scene features, while Parking-Distill improves the learning quality of small and distorted targets through teacher soft labels and a difficult target compensation mechanism, making it easier for the model to converge to a parameter distribution suitable for the parking perception task. Figure 7 Data shows that this algorithm not only improves the final detection accuracy, but also improves the convergence speed and stability of the training process, and can reduce the problem of insufficient learning of the model on difficult targets such as low obstacles, curbs, and speed bumps.

[0098] Figure 8 A categorized comparison of parking safety assessments (APs) was conducted for six types of parking targets: low obstacles, people, vehicles, curbs, speed bumps, and pillars. The results show that the improvements in this algorithm are mainly concentrated on key and challenging parking safety targets, including low obstacles, curbs, speed bumps, and pillars. Specifically, the accuracy rate for low obstacles improved from 61.2% in the basic YOLOv10s algorithm to 80.6% in the complete algorithm; curbs from 54.3% to 73.8%; speed bumps from 58.9% to 77.5%; and pillars from 63.8% to 74.4%.

[0099] In contrast, people and vehicles have more pronounced appearance textures and larger target sizes, resulting in a relatively stable improvement, although the overall algorithm still maintains a steady increase. This indicates that the algorithm's enhancement mechanism is highly targeted: without compromising general target detection capabilities, it focuses on strengthening the detection of low, slender, and close-range distorted targets in parking scenarios. Figure 8 Data shows that this algorithm has the characteristics of "enhancing difficult targets and maintaining the stability of common targets", which can better meet the detection needs of parking environment perception for safety-critical targets.

[0100] Figure 9Different models were comprehensively evaluated from three dimensions: accuracy, speed, and parameter count. The horizontal axis represents inference speed (FPS), tested on OrinNx hardware; higher inference speed is better. The vertical axis represents mAP50 (%), with higher values ​​being better. The bubble position indicates the model's mAP50 and inference speed, and the bubble size indicates the number of parameters. The results show that YOLOv8s has a large number of parameters and lower accuracy; YOLOv10s has advantages in terms of lightweight design and real-time performance; after adding Parking-LoRA, the model's parameter count only increased from 7.2M to approximately 7.34M, while the inference speed remained at approximately 71.5 FPS, but mAP50 was significantly improved. Further adding Parking-Distill, which primarily functions as a distillation constraint during training with zero additional overhead during inference, resulted in the complete algorithm achieving the highest accuracy while maintaining 71.5 FPS.

[0101] These results demonstrate that our algorithm does not rely on large models, complex attention mechanisms, or expensive post-processing, but rather performs structural enhancements within the lightweight YOLOv10s framework. The parameter increment for Parking-LoRA is approximately 1.9%, and the Parking-Distill inference stage incurs zero additional overhead, allowing the model to still meet the real-time, low-power, and low-latency requirements of automotive edge computing platforms. Figure 9 The data shows that this algorithm achieves a good balance between accuracy improvement and deployment efficiency, and is suitable for deployment on automotive embedded platforms such as OrinNX for real-time low-speed parking environment perception.

[0102] Figure 10 The overall capabilities of different models across dimensions such as low-profile recall, curbstone AP, speed bump AP, near-field localization, low false negatives, and real-time performance are comprehensively presented in the form of normalized radar charts. The basic YOLOv10s performs well in real-time performance, but has shortcomings in low-profile recall, slender target detection, and near-field localization. After adding Parking-LoRA, the dimensions of low-profile recall, curbstone AP, and speed bump AP are significantly expanded, indicating that this module has a significant feature enhancement effect on low-profile, slender targets. The complete algorithm further expands in the dimensions of near-field localization and low false negatives, demonstrating the compensating effect of Parking-Distill on near-field distortion and confidence in difficult targets.

[0103] The radar chart shows that the complete algorithm outperforms the base model in several key capabilities, particularly in parking safety-related dimensions. Meanwhile, real-time performance remains largely unchanged, indicating that the algorithm enhancements did not significantly sacrifice inference speed. Figure 10 Data shows that this algorithm has a more balanced overall parking perception capability, achieving a better balance between low target recall, slender target recognition, near-range positioning stability, and real-time deployment.

[0104] In summary, simulation results demonstrate that the proposed YOLOv10s+Parking-LoRA+Parking-Distill parking environment perception algorithm exhibits higher detection accuracy and lower safety risk indicators for key parking safety targets such as low-lying obstacles, curbs, speed bumps, and pillars. Compared to the basic YOLOv10s, this algorithm enhances the feature representation of low-lying, slender targets through Parking-LoRA and corrects near-range distortion and compensates for low-confidence, difficult targets through Parking-Distill. This significantly improves the recall rate of low-lying obstacles, the AP of curbs, the AP of speed bumps, and the F1-Score of curbs, while effectively reducing near-range localization errors, the false negative rate of low-lying obstacles, and the false positive rate of pillars / vehicle corners. Regarding model deployment, this algorithm introduces only a small number of parameter increments, and there is no additional computational overhead in the Parking-Distill inference stage, maintaining overall real-time inference capabilities. It is suitable for automatic parking and low-speed near-field obstacle avoidance environment perception tasks at vehicle speeds below 30 km / h.

Claims

1. An autonomous driving environment perception method, characterized in that... It includes the following steps: S1: Construct a scale-aware dynamic strip convolution module; The scale-aware dynamic bar convolution module includes: a lightweight feature morphological statistics layer and a dynamic one-dimensional decomposition convolution layer connected in sequence. The lightweight feature morphological statistics layer processes the input scale feature map X. s Perform global statistics and output the dynamic kernel length coefficient k. s And feed it into the dynamic one-dimensional decomposition convolution; k s =F stat (X s ); In the formula, F stat For dynamic kernel length coefficient k s A lightweight feature morphological statistics layer; s is the scale corresponding to the input image; The dynamic one-dimensional decomposition convolution is based on the dynamic kernel length coefficient k. s Construct sequential horizontal and vertical depthwise convolutions, and add orientation-constrained offsets to output the extracted bar features F. long,s Dynamic one-dimensional decomposition convolution includes the following operations: offset= F offset (X s ); F hori,s =DWConv 1×k_s,offset (X s ); F long,s =DWConv k_s×1,offset (F hori,s ); In the formula, DWConv is a depthwise separable convolution, 1×k s Represents adaptive horizontal bar convolution, k s ×1 represents adaptive vertical stripe convolution; offset is the orientation-constrained offset; F hori,s To calculate the intermediate value; F offset A lightweight feature mapping layer oriented towards orientation-constrained offsets; S2: Construct a dynamically dilated, dilated convolutional module; The dynamic dilation rate dilated convolution module is designed for input feature X. s Adaptive selection of expansion rate d s It dynamically matches the context range of obstacles of different sizes and outputs the context features F of low obstacles. small,s ; d s =F scale (X s ); F small,s =DWConv 3×3,d (X s ); In the formula, d takes the value d s ;F scale For the expansion rate d s A lightweight feature-scale statistical layer; S3: Construct a branch attention interaction fusion module; The branch attention interaction fusion module achieves bidirectional complementary enhancement of features of slender targets and features of low-lying targets through a lightweight channel attention interaction mechanism, and outputs the fused feature F. fuse,s ; To long =CA(F long,s )⊙F small,s ; To small =CA(F small,s )⊙F long,s ; F fuse,s = F long,s + To long + F small,s + To small ; In the formula, CA(·) is the lightweight channel attention module, and ⊙ represents element-wise multiplication; Att long Features extracted for slender targets; Att small Features extracted for low-lying targets; S4: Constructing a joint gating system G s The joint gating G s Based on pre-generated dual prior parameters W dist and W pos Enhance distorted regions and suppress invalid backgrounds; G s =W dist ⊙W pos ; In the formula, W dist For the distortion weights acquired from the input image; W pos For perspective position weights acquired from the input image; S5: Construct the residual module; the residual module uses a residual structure to ensure gradient flow, and applies gated weights to the fused features to obtain the enhanced features F. out,s ; F out,s =X s +Gs⊙F fuse,s ; S5: Construct the target focusing module; The target focusing module includes: a scale-aware dynamic strip convolution module, a dynamic dilation rate dilated convolution module, a branch attention interaction fusion module, and a joint gating G... s and the residual module; Input feature X s The features are fed into the scale-aware dynamic strip convolution module and the dynamic dilation rate dilated convolution module, which are set in parallel. The features output by the two modules are simultaneously fed into the branch attention interaction fusion module for complementary fusion, and then sequentially fed into the joint gating G. s After processing in the residual module, the enhanced feature F is output. out,s ; S6: Construct an environment perception model based on a lightweight YOLO model; The environmental perception model includes: a backbone network, a neck network, a target focusing module, and a detection head; The backbone network, neck network, and detection head are implemented based on the standard structure of the YOLO model. A target focusing module is inserted between each detection head and the neck network. The target focusing module enhances the feature data output by the neck network, and then the enhanced feature F is processed. out,s It is then fed into the corresponding detection head for subsequent target detection.

2. The autonomous driving environment perception method according to claim 1, characterized in that: In step S1, the dynamic kernel length coefficient k is... s Lightweight feature morphological statistical layer F stat and a lightweight feature mapping layer F with orientation-constrained offset offset The structure is the same, including: a globally average pooled pool and two fully connected pools connected in sequence.

3. The autonomous driving environment perception method according to claim 1, characterized in that: In step S2, the expansion rate d is... s Lightweight feature-scale statistical layer F scale It includes: a globally average pooling GAP layer connected in sequence, two fully connected layers connected in sequence, and three parallel classification layers; The input feature map is fed into a global average pooling (GAP) layer to extract features G representing the target scale and global density statistics of the entire feature map. Two fully connected layers are sequentially connected, and feature G is fed into the first fully connected layer FC1 to refine and compress the scale features F1. F1 = σ(FC1(G)); where σ is the ReLU activation function; The compressed scale feature F1 is then fed into the second fully connected layer FC2, which consists of three neurons and is mapped to three scores S corresponding to the three dilation rates. S = FC2 (F1); Three classification layers are connected in parallel, and output scores S corresponding to low-dense targets, medium targets, and low-sparse targets, respectively. In the classification layer, based on a pre-set discrete inflation rate, the neuron index with the highest score is selected using argmax and mapped to the final inflation rate d. s :d s =argmax(S)+1.

4. The autonomous driving environment perception method according to claim 1, characterized in that: In step S4, the distortion weight W dist The calculation method is as follows: W dist =σ(λ1·r(x,y)); Perspective position weight W pos The calculation method is as follows: W pos =σ(λ2·y / H); In the formula, r(x,y) represents the normalized radial distance corresponding to pixel (x,y); r(x,y)=sqrt((x-W / 2) 2 +(y-H / 2) 2 ) / (sqrt(W 2 +H 2 ) / 2); sqrt() is the arithmetic square root function; σ is the sigmoid activation function; λ1 and λ2 are learnable hyperparameters; H is the scale feature map X. s The image height, W, is the scale feature map X. s The width.

5. The autonomous driving environment perception method according to claim 1, characterized in that: The environmental perception model also includes: a correction module; The correction module is located after the output of the detection head and before the non-maximum suppression (NMS) process. The correction module corrects the prediction result output by the detection head, performs subsequent NMS processing on the corrected prediction result, and then outputs the final result. The correction module includes: a near-range distortion sensing correction module, a small target soft distillation guidance module, a difficult target confidence compensation module, and a correction output module connected in sequence. The near-field distortion perception and correction module addresses the significant radial distortion produced by the parking wide-angle lens at close range by establishing a physical compensation model; it also establishes a near-field distortion perception and correction mapping function T(·) to correct the bounding box coordinate offset and outputs the corrected bounding box (bbox). i corr ; The difficult target confidence compensation module designs a confidence compensation mechanism to identify and correct such targets for low-confidence difficult samples, and outputs the compensated target confidence score (conf). i corr ; To address the issue of naturally low confidence levels for low-lying obstacles, the small-target soft distillation guidance module uses high-quality teacher signals to perform soft distillation constraint optimization on the target category, outputting a corrected category prediction vector cls. i corr ; The corrected output module constructs the corrected prediction result Output from the set of final perception results corresponding to all targets within a frame. crre The result is then fed into a subsequent module for non-maximum suppression (NMS) calculation. Output crre ={Out1,Out2,..., Out i ,…,Out M }; The corrected output for the i-th target is: Out i =(argmax(cls i corr ), bbox i corr , conf i corr ); In the formula, M is the total number of targets detected in the input image of this frame; argmax(cls i corr The expression indicates that the index of the maximum value of the corrected category vector is used as the final target category ID.

6. The autonomous driving environment perception method according to claim 5, characterized in that: In the near-field distortion perception and correction module, a near-field distortion perception and correction mapping function T(·) is constructed to correct the box coordinate offset for radial distortion of the parking near-field camera; bbox i corr =T(bbox i ,D i ); bbox i corr D represents the corrected target bounding box corresponding to the i-th target; i Let be the distance between the i-th target and the vehicle; The following operations are performed in the near-range distortion sensing correction mapping function T(·): a1: Determine D i Does it fall within the calibration range? If the error does not fall within the calibration range, no calibration operation will be performed. If it falls within the correction range, then perform the following steps; a2: Radial distance calculation: r = sqrt((cx - W / 2)) 2 +(cy-H / 2) 2 ) / (sqrt(W 2 +H 2 ) / 2); In the formula, r is the normalized radial distance; sqrt() is the arithmetic square root function; cx and cy are the pixel coordinates of the center point of the current target bounding box; W and H are the width and height of the input image; a3: Calculation of radial distortion stretching factor: radial_factor=1+k1*r 2 +k2*r 4 ; k1 and k2 are the camera radial distortion coefficients; radial_factor is the radial distortion stretching coefficient; a4: Bounding box width and height compensation: bbox_w corrected =box_w / radial_factor. bbox_h corrected =bbox_h / radial_factor; In the formula, bbox_w and bbox_h are the width and height of the original bounding box output by the model prediction head; bbox_w corrected To compensate for the distortion of bbox_w, bbox_h corrected This is the height of the bounding box after distortion compensation for bbox_h; a5: Bounding box center point position compensation: c x_corrected =cx·(1+△ cx ·r); c y_corrected =cy·(1+△ cy ·r); In the formula, △ cx and △ cy It is the target bounding box center point offset compensation coefficient; c x_corrected and c y_corrected The coordinates of the center point of the frame after compensation; a6: Adjust the corrected bounding box width (bbox_w) corrected Corrected bounding box height bbox_h corrected and the coordinates of the center point of the bounding box (c x_corrected ,c y_corrected ) constitute bbox i corr The corresponding values.

7. The autonomous driving environment perception method according to claim 5, characterized in that: The small target soft distillation guidance module includes the following operations: b1: Construct a small target mask to filter the distillation target; small_mask=(bbox_area <Sdt); In the formula, bbox_area is the pixel area of ​​the bounding box of the current i-th target, bbox_area=w i ×h i Sdt is the threshold for small target determination; small_mask is a boolean mask for small targets; b2: Construct the KL divergence distillation loss formula; L ~ kd =KLDivLoss(F.log_softmax((cls_student) / T), F.softmax((cls_teacher) / T))×T 2 ; In the formula, L ~ kd KLDivLoss is the KL divergence distillation loss; KLDivLoss is the function for calculating the KL divergence loss. cls_student represents the student model, and cls_teacher represents the soft label for the teacher model; F.log_softmax is the logarithmic softmax activation function, and F.softmax is the standard softmax activation function; T is the distillation temperature; T 2 This is a temperature coefficient compensation term, used to offset the effect of distillation temperature on the loss amplitude; b3: Mask Filtering: Distillation loss is filtered by a small target mask, retaining only the loss of small targets, and the average value of the retained small target losses is calculated to obtain L. kd ; L kd =(∑L ~ kd,i ·small_mask) / N; In the formula, N is the total number of samples in a single batch; L ~ kd,i This represents the divergence distillation loss corresponding to the i-th target within this batch; b4: Construct the total loss function L of the model total ; L total =L det +DLW×L kd ; In the formula, L det DLW represents the native detection loss output by the detection head of the environmental perception model; DLW represents the distillation loss weights. b5: Based on soft label fusion, the category prediction results are corrected; The training phase integrates teacher model knowledge to optimize category output: cls i corr = softmax( cls i + α·(cls teacher - cls i ) ); In the formula, α is the distillation-fusion weight; cls i corr This is the corrected category prediction vector, used for classification output; cls i This is the original prediction vector for the i-th category output by the detection head of the environmental perception model.

8. The autonomous driving environment perception method according to claim 5, characterized in that: The confidence compensation module for difficult targets includes the following operations: c1: Construct the aspect ratio prior factor ar_prior; ar_prior=σ(w i / (h i +10 -6 ) - Apt); In the formula, w i and h i The width and height of the original bounding box corresponding to the i-th target; 10 -6 To prevent the division into zero minimum values; σ is the Sigmoid activation function; Apt is the preset aspect ratio prior threshold; c2: Construct the location prior factor pos_prior; Calculate the normalized ordinate c' of the center point pixel of the current target bounding box. y :c' y =c y / H; where H is the height of the input image; Compare c'y with the preset position prior threshold Lpt; When c' y When Lpt is greater than 1, pos_prior = 1.0; When c' y When Lpt is less than or equal to Lpt, pos_prior = 0.5; c3: Construct the comprehensive compensation coefficient △conf; △conf i =△×ar_prior×pos_prior; In the formula, △ represents the preset maximum compensation range; △conf i This represents the comprehensive compensation coefficient corresponding to the i-th objective; c4: For the i-th detection target, determine whether the following four conditions are met simultaneously; if they are met simultaneously, execute c5 to perform confidence compensation; otherwise, do not perform confidence compensation for difficult targets, and directly proceed to the next step; Condition 1: The target category belongs to the hard target set HARD_CLASSES; The set of difficult targets, HARD_CLASSES, includes: low obstacles, curbs, and speed bumps; Condition 2: Initial confidence level (conf) i Condition met: Confidence lower limit threshold θ low <conf i <Confidence threshold θ high ; Condition 3: The target is elongated; the method for determining the elongated shape is that the aspect ratio of the target bounding box is greater than the prior threshold Apt for the aspect ratio. Condition 4: The target is located in the near-ground region of the image: the normalized ordinate c' of the center pixel of the current target bounding box. y Location prior threshold Lpt; c5: Compensated confidence level (conf) corr The calculation method is as follows: conf i corr = max(conf i + △conf i , θ low ), and conf i corr ≤ 0.9; In the formula, θ low The preset confidence level is the safety lower bound; conf i corr Let be the confidence level after final compensation for the i-th objective.

Images