A detection device mounted on a vehicle side and an unstructured road edge detection method

Abstract

The application discloses a detection device mounted on a vehicle side and an unstructured road edge detection method, and relates to the technical field of video recognition. Step one: an original image enters a ResNet50 backbone network, layer-by-layer convolution and down-sampling are performed, and features of different levels are respectively output; step two: deep layer features output by the backbone network are input into a Deformable.ASPP module, and multi-scale enhanced features of a fixed channel number are output; step three: high layer features of the ASPP are sent into an edge enhancement module, edge features are extracted, the edge features are spliced and fused with ASPP output in a channel dimension, and multi-level features are extracted by the ResNet50 backbone network, wherein shallow layer features contain rich edge and texture information, and deep layer features contain global semantic information. Then, multi-scale feature extraction is performed through the Deformable ASPP module, deformable convolution is used to adaptively capture target regions of different scales and deformations, and the weight of each branch is dynamically adjusted through a controller.

Description

Technical Field

[0001] This invention relates to the field of video recognition technology, specifically to a detection device mounted on the side of a vehicle and a method for detecting unstructured road edges. Background Technology

[0002] Road edge detection is a crucial foundation for automated driving, and its accuracy directly impacts vehicle path planning and driving safety. Compared to structured roads such as highways and urban roads, unstructured roads often lack clear lane lines and edge markings, have complex road surface materials, blurred boundaries, and are easily affected by external environmental interference, making detection more challenging.

[0003] Early detection methods primarily relied on traditional image processing techniques, such as edge operator detection, thresholding, and geometric fitting. While simple to implement, these methods had limited performance in complex environments. Subsequent machine learning methods improved robustness by utilizing handcrafted features and statistical modeling, but their generalization ability remained insufficient. With the development of deep learning, semantic segmentation models based on convolutional neural networks have gradually become mainstream. They effectively improve edge detection performance in unstructured scenes through multi-scale feature fusion and contextual modeling. However, unstructured road edge detection still faces many challenges, including scene diversity, edge ambiguity, the influence of lighting and weather, and real-time requirements. Summary of the Invention

[0004] The purpose of this invention is to provide a detection device mounted on the side of a vehicle and a method for detecting unstructured road edges, so as to solve the problems mentioned in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an unstructured road edge detection method, comprising the following steps: Step 1: The original image is fed into the ResNet50 backbone network, and convolution and downsampling are performed layer by layer to output features at different levels. Step 2: Input the deep features output by the backbone network into the Deformable.ASPP module to output multi-scale augmented features with a fixed number of channels; Step 3: Feed the high-level features of ASPP into the edge enhancement module to extract edge features, and then stitch and fuse the edge features with the ASPP output in the channel dimension. Step 4: The edge-enhanced features are input together with features from different levels into a multi-level fusion decoder to unify the features from different levels to the same channel dimension and then splice them together. Step 5: Feed the decoded and fused features into the classifier, sample them to the same resolution as the input image to form a class response map, and then use bilinear interpolation.

[0006] Preferably, in step one: after the acquired original image enters the ResNet50 backbone network, during the process of layer-by-layer convolution and downsampling, the network outputs features at different levels: shallow features (from layer 1) retain detailed information such as edges and textures; mid-level features (from layers 2 and 3) take into account certain semantic and structural information; and deep features (from layer 4) contain global context and high-level semantics.

[0007] Preferably, in step two: after obtaining the deep features output by the backbone network, they are input into the Deformable.ASPP module, which consists of a 1×1 convolutional branch, multiple deformable convolutional branches and a global pooling branch. The deformable convolutional branches can adaptively capture irregular shapes and complex boundaries by learning the sampling position offset, while the global pooling branch introduces global context information. At the same time, the module is designed with a dynamic weight controller to adaptively adjust the response intensity of each branch according to the input features. Finally, all branch features are concatenated in the channel dimension and fused by convolution to output multi-scale enhanced features with a fixed number of channels. Among them, ASPP runs a dynamic hole rate mechanism. To address the problem of significant differences in the scale of unstructured roads, it proposes to introduce a dynamic hole rate into the ASPP module. The dilation coefficient of the dilated convolution is adaptively adjusted according to the content of the input feature map, avoiding the loss of semantic information caused by a fixed hole rate, thereby enhancing the model's ability to perceive road structures at different scales. The ASPP running branch weighting strategy suppresses redundant features and highlights key scale information, enabling the model to have a stronger ability to express multi-scale features in complex backgrounds.

[0008] Preferably, in step three: the high-level features obtained by ASPP are fed into the edge enhancement module. The edge enhancement module first uses a 3×3 convolution (similar to Sobel convolution) to extract potential edge features, and then concatenates the edge features with the ASPP output in the channel dimension. The concatenated features are fused through an additional convolution block, thereby enhancing semantic information while highlighting the feature expression of the edge region. The edge enhancement module strengthens the representation of road boundary features, alleviates the problems of boundary blurring and breakage in unstructured road segmentation, and significantly improves the accuracy and continuity of boundary detection.

[0009] Preferably, in step four: the edge-enhanced features are input together with the features from layer1, layer2 and layer3 into the multi-level fusion decoder. The decoder first unifies the features from different layers to the same channel dimension through a series of 1×1 convolutions, and then uses bilinear interpolation to upsample them to the same spatial resolution as layer1. Next, all features are concatenated in the channel dimension and then fused through two layers of 3×3 convolutions to achieve complementarity between shallow details and deep semantics. The multi-level fusion decoder described in this invention differs from the single high-low layer feature fusion of the traditional DeepLabv3+. The multi-level feature fusion decoding structure designed in this invention ensures the consistency of the feature map before fusion through projection layers and upsampling, and fuses deep semantic information with shallow fine-grained features, which not only maintains global consistency, but also enhances the ability to restore details of local boundaries and textures.

[0010] Preferably, in step five: the decoded and fused features are fed into a classifier, a class response map is generated using 1×1 convolution, the number of channels is equal to the number of classes in the segmentation task, and finally, bilinear interpolation is used to upsample the result to the same resolution as the input image to obtain a segmentation prediction map.

[0011] A detection device mounted on the side of a vehicle includes a vehicle that provides mobility and a camera device for capturing raw images. A support assembly is installed on the top of the vehicle. The support assembly drives two internally threaded parts, each of which is internally threaded with an externally threaded part. A rubber pad is fixedly connected to one side of each externally threaded part, and a hexagonal rod and a mounting bracket are provided between the other side of the externally threaded part and the camera device. A pressure control assembly is provided between the two externally threaded parts.

[0012] Preferably, the support assembly includes a robotic arm rotatably connected to the top of the vehicle, a pneumatic cylinder fixedly connected to the top of the robotic arm, a mounting frame one fixedly connected to the piston end of the pneumatic cylinder, a forward and reverse motor fixedly connected to one side of the mounting frame one, a drive shaft fixedly connected to the output end of the forward and reverse motor, a mounting frame three fixedly connected to one end of the drive shaft, a mounting frame four fixedly connected to the mounting frame three, and a mounting frame five fixedly connected to one side inside the mounting frame four. Two telescopic rods are fixedly connected to the top of the robotic arm, with the piston end of each telescopic rod fixedly connected to the mounting frame one. A flat needle roller bearing is fixedly connected to one side of the mounting frame one, and two connecting rods are fixedly connected between the flat needle roller bearing and the mounting frame three. A brake is fitted on the outer side of the shaft. Two mounting brackets are fixedly connected between the brake and mounting bracket one. An operating rod and two support shafts are rotatably connected between mounting bracket five and mounting bracket four. A gear three is fixedly installed on the outer side of the operating rod. Gears two are fixedly installed on the outer sides of both support shafts two. Both gears two mesh with gear three. A friction element one is provided on one side of each gear two. The friction element one is fixedly connected to mounting bracket four. Two internal threaded parts are rotatably passed through mounting bracket five. Gear one is fixedly connected to the outer side of each internal threaded part. Gear one meshes with the adjacent gear two. A support shaft one is rotatably connected between the internal threaded part and mounting bracket four.

[0013] Preferably, the external threaded component has multiple filler plates fixedly connected inside, the hexagonal rod is slidably inserted inside the mounting frame six, one end of the hexagonal rod is fixedly connected to the external threaded component, and the mounting frame six is ​​fixedly connected to the camera equipment.

[0014] Preferably, the pressure control assembly includes a sealed box fixedly connected to one side of the camera device, a moving plate disposed inside the sealed box, two air-storing bladders fixedly connected to the top of the moving plate, and a lead screw rotatably passing through the moving plate. The lead screw is threadedly connected to the moving plate, and the top end of the lead screw extends to the top of the sealed box. A cross-shaped button is fixedly connected to the top end of the lead screw. Two friction elements are fixedly installed on the outside of the lead screw. Friction elements three and four are respectively sleeved on the outside of the two friction elements two. Friction elements three and four are both fixedly installed inside the sealed box. Two guide vertical rods pass through the moving plate and are fixedly installed inside the sealed box. The top end of the air-storing bladder is fixedly connected to the top of the inner cavity of the sealed box. A flexible hose is disposed on the top of the air-storing bladder and is fixedly installed on the top of the sealed box. One end of the flexible hose is fixedly connected to an adjacent external threaded component. An air guide groove is disposed on one side of the flexible hose and is located behind the external threaded component.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In this application, a ResNet50 backbone network extracts multi-level features, where shallow features contain rich edge and texture information, and deep features contain global semantic information. Then, a Deformable ASPP module is used for multi-scale feature extraction, utilizing deformable convolutions to adaptively capture target regions of different scales and deformations, while a controller dynamically adjusts the weights of each branch. Next, an edge enhancement module is used to extract significant edge features using Sobel-like convolutions, which are then fused with the ASPP output features to improve the expressive power of boundary regions. The enhanced features, along with features from intermediate and shallow layers, are input into a multi-level fusion decoder, where they are concatenated and fused after unifying the number of channels and scale, achieving an effective combination of details and semantics. Finally, a classifier generates semantic segmentation results, and upsampling is used to restore the resolution to the input image, resulting in accurate pixel-level predictions.

[0016] 2. When using this application, push the rubber pad fixedly connected to one side of the external threaded component to adhere to the clean side wall of the vehicle. Because the hexagonal rod is slidably inserted into the mounting bracket six, the two rubber pads can be on different planes. The two deformable rubber pads can simultaneously adhere to the outer wall of the vehicle. Apply a certain thrust to the camera device to deform the two rubber pads and seal the inside of the external threaded component. Then, operate the cross knob to rotate clockwise. The screw rotates clockwise to drive the moving plate to move horizontally downward. The downward movement of the moving plate causes the air storage bag to unfold and draw air from the inside of the external threaded component through the hose. The inside of the external threaded component enters a negative pressure state. The multiple filling plates fixedly connected inside the external threaded component reduce the amount of air stored inside the external threaded component and increase the speed of negative pressure rise inside the external threaded component. As the moving plate moves downward, the negative pressure inside the external threaded component strengthens and adheres to the outer wall of the vehicle. Finally, the camera device is installed and suspended on the outside of the vehicle. It can quickly and conveniently replace the vehicle carrying the camera device, adapt to road data collection work in different environments, and increase the flexibility of using the detection equipment mounted on the side of the vehicle. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the unstructured road edge detection method of the present invention; Figure 2 The experimental loss curve of the unstructured road edge detection method is shown. Figure 3 The experimental performance curves for unstructured road edge detection methods are shown. Figure 4 A comparison of the processing effects of unstructured road edge detection methods; Figure 5 This is a schematic diagram of the structure of the robotic arm of the present invention; Figure 6 This is a schematic diagram of the structure of the mounting bracket of the present invention; Figure 7This is a schematic diagram of the structure of mounting bracket three of the present invention; Figure 8 This is a schematic diagram of the separation structure of the drive shaft and mounting bracket three of the present invention; Figure 9 This is a schematic diagram of the separation structure of mounting bracket four and mounting bracket five of the present invention; Figure 10 This is a schematic diagram of the structure of mounting bracket five of the present invention; Figure 11 This is a schematic diagram of the structure of the sealing box of the present invention; Figure 12 This is a cross-sectional view of the sealing box of the present invention; Figure 13 This is a schematic diagram of the structure of the motion plate of the present invention.

[0018] Numbered in the diagram: 1. Vehicle; 2. Robotic arm; 3. Telescopic rod; 4. Pneumatic cylinder; 5. Mounting bracket one; 6. Forward and reverse motor; 7. Drive shaft; 8. Mounting bracket two; 9. Brake; 10. Mounting bracket three; 11. Flat needle roller bearing; 12. Connecting rod; 13. Mounting bracket four; 14. Support shaft one; 15. Internal threaded component; 16. Gear one; 17. Friction component one; 18. Support shaft two; 19. Gear two; 20. 21. Gear 3; 22. Mounting bracket 5; 23. External threaded part; 24. Rubber pad; 25. Filler plate; 26. Hexagonal rod; 27. Mounting bracket 6; 28. Camera equipment; 29. ​​Sealing box; 30. Moving plate; 31. Air storage bag; 32. Hose; 33. Air guide groove; 34. Guide vertical rod; 35. Lead screw; 36. Friction component 2; 37. Friction component 3; 38. Friction component 4; 39. Cross button; 30. Operating lever. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figures 1-4 As shown, the present invention provides a technical solution for an unstructured road edge detection method, comprising the following steps: Step 1: The input raw image first enters the ResNet50 backbone network. During the layer-by-layer convolution and downsampling process, the network outputs features at different levels: shallow features (from layer 1) preserve details such as edges and textures; mid-level features (from layers 2 and 3) take into account certain semantic and structural information; deep features (from layer 4) contain global context and high-level semantics. Step 2: After obtaining the deep features output by the backbone network, they are input into the Deformable ASPP module. This module consists of a 1×1 convolutional branch, multiple deformable convolutional branches, and a global pooling branch. The deformable convolutional branches can adaptively capture irregular shapes and complex boundaries by learning the sampling position offset, while the global pooling branch introduces global contextual information. Simultaneously, the module employs a dynamic weight controller to adaptively adjust the response intensity of each branch based on the input features. Finally, all branch features are concatenated along the channel dimension and fused through convolution to output multi-scale enhanced features with a fixed number of channels. Its ASPP dynamic dilatation rate mechanism addresses the problem of significant differences in the scale of unstructured roads. It proposes to introduce a dynamic dilatation rate into the ASPP module, which adaptively adjusts the dilation coefficient of the dilated convolution based on the content of the input feature map, avoiding the loss of semantic information caused by a fixed dilatation rate, thereby enhancing the model's ability to perceive road structures at different scales.

[0021] Its ASPP branch weighting strategy suppresses redundant features and highlights key scale information, enabling the model to have a stronger ability to express multi-scale features in complex backgrounds.

[0022] Step 3: The high-level features from ASPP are fed into the edge enhancement module. This module first uses a 3×3 convolution (similar to a Sobel convolution) to extract potential edge features, and then concatenates the edge features with the ASPP output along the channel dimension. The concatenated features are then fused through an additional convolutional block, thereby enhancing semantic information while highlighting the feature representation of edge regions. Its edge enhancement module strengthens the representation of road boundary features, alleviates the problems of boundary ambiguity and breakage in unstructured road segmentation, and significantly improves the accuracy and continuity of boundary detection.

[0023] Step 4: The edge-enhanced features, along with features from layers 1, 2, and 3, are input into the multi-level fusion decoder. The decoder first unifies the features from different layers to the same channel dimension using a series of 1×1 convolutions, then upsamples them to the same spatial resolution as layer 1 using bilinear interpolation. Next, all features are concatenated along the channel dimension and fused using two 3×3 convolutions to achieve complementarity between shallow details and deep semantics. Unlike the single high-low layer feature fusion of the traditional DeepLabv3+, the multi-level feature fusion decoding structure designed in this invention ensures the consistency of the feature map before fusion through projection layer and upsampling, and fuses deep semantic information with shallow fine-grained features, which not only maintains global consistency, but also enhances the ability to restore details of local boundaries and textures.

[0024] Step 5: The decoded and fused features are fed into the classifier, where a 1×1 convolution is used to generate a class response map, with the number of channels equal to the number of classes in the segmentation task. Finally, bilinear interpolation is used to upsample the result to the same resolution as the input image, resulting in a segmentation prediction map.

[0025] like Figures 5-13 The detection equipment shown, mounted on the side of the vehicle, includes a vehicle 1 that provides mobility and a camera 27 that captures raw images. A support assembly is mounted on the top of the vehicle 1. The support assembly drives two internally threaded parts 15. Both internally threaded parts 15 are internally threaded with externally threaded parts 22. A rubber pad 23 is fixedly connected to one side of the externally threaded parts 22. A hexagonal rod 25 and a mounting bracket 26 are provided between the other side of the externally threaded parts 22 and the camera 27. A pressure control assembly is provided between the two externally threaded parts 22.

[0026] The working principle of the detection equipment mounted on the side of the vehicle is as follows: In the support assembly, the robotic arm 2 is fixedly installed on the top of the vehicle 1. The piston end of the pneumatic cylinder 4 is fixedly connected to the top of the robotic arm 2, and the mounting bracket 5 is fixedly connected to the top of the pneumatic cylinder 4. Two telescopic rods 3 are fixedly connected to the top of the robotic arm 2, and the piston ends of the telescopic rods 3 are fixedly connected to the mounting bracket 5. With the cooperation of the telescopic rods 3 and the pneumatic cylinder 4, the relative distance between the mounting bracket 5 and the robotic arm 2 is adjusted, and the movement of the camera device 27 supported by the support assembly is controlled, so that the camera device 27 can be suspended on the side of the vehicle 1. The operation of the robotic arm 2 drives the telescopic rods 3 and the pneumatic cylinder 4 to rotate, and the mounting bracket 5 to rotate, so that the camera device 27 can move to various positions on the outside of the vehicle 1, ensuring the flexibility of the camera device 27 and enabling the camera device 27 to effectively avoid obstacles.

[0027] A drive shaft 7 is fixedly connected to the output end of a forward / reverse motor 6, which is fixedly connected to one side of mounting bracket 15. A mounting bracket 3 10, which is fixedly connected to one end of the drive shaft 7, has a mounting bracket 4 13 fixedly connected to it. The forward / reverse motor 6 controls the operation of the forward / reverse motor 6, driving the mounting bracket 3 10 to rotate. Two connecting rods 12 are fixedly connected between a flat needle roller bearing 11, which is fixedly connected to one side of mounting bracket 15, and the mounting bracket 3 10. These connecting rods 12 support the mounting bracket 3 10, reducing the workload of the forward / reverse motor 6. A brake 9 is fitted on the outside of the drive shaft 7. Two mounting brackets 2 8 are fixedly connected to the brake 9 and mounting bracket 15. The brake 9 controls the operation of the drive shaft 7, limiting the movement of mounting bracket 3 10 and the rotation of mounting bracket 4 13. A mounting bracket 5 21 and an internally threaded component 1 are fixedly connected inside mounting bracket 4 13. 5. The device is mounted on the mounting frame 21. When the mounting frame 13 rotates, the mounting frame 21 drives two internally threaded parts 15 to rotate around the drive shaft 7. Both internally threaded parts 15 are internally threaded with externally threaded parts 22. The hexagonal rod 25, which is fixedly connected to the externally threaded parts 22, is slidably inserted into the mounting frame 26. The mounting frame 26 is fixedly connected to the camera device 27. At this time, under the action of the two externally threaded parts 22, the two hexagonal rods 25, and the two mounting frames 26, the camera device 27 rotates. By adjusting the pitch angle of the camera device 27 and controlling the operation of the support components, the pitch angle of the camera device 27 can be adjusted so that the shooting path of the camera device 27 forms different angles with the ground. This ensures that the camera device 27 can shoot the road surface flexibly and efficiently, and allows the camera device 27 to capture effective original images.

[0028] Additionally, an operating lever 39 and two support shafts 28 are rotatably connected to a mounting bracket 5 21 fixedly connected to one side inside the mounting bracket 4 13. A gear 3 20 is fixedly installed on the outside of the operating lever 39, and the gear 3 20 can rotate inside the mounting bracket 4 13. Gears 2 19 are fixedly installed on the outside of each of the two support shafts 2 18, and the two gears 2 19 can rotate. Both gears 2 19 mesh with gears 3 20. Two internal threaded parts 15 are rotatably mounted on the mounting bracket 5 21. A gear 1 16 fixedly connected to the outside of the internal threaded part 15 meshes with the adjacent gear 2 19. A support shaft 1 14 is rotatably connected between the internal threaded part 15 and the mounting bracket 4 13. The support shaft 1 14 increases the stability of the internal threaded part 15. A friction element 1 17 is fixedly connected to the mounting bracket 4 13 on one side of the gear 2 19. There is a certain friction between the friction element 1 17 and the gear 2 19, so that the operating lever 39 will not rotate automatically due to vibration, thus ensuring the stability of the camera equipment 27. Therefore, after holding the camera device 27 and limiting its position, the operating lever 39 is located on the outer side of the mounting bracket 13. The operating lever 39 is rotated through gear 3 20, two gears 2 19 and two gears 1 16, causing the two internal threaded parts 15 to rotate. The internal threaded parts 15 rotate away from the outer side of the limited external threaded parts 22, thus completing the disassembly of the camera device 27 from the support assembly. The camera device 27 can be quickly removed from the vehicle 1 and dropped, allowing the user to hold and use the camera device 27, making the camera device 27 adaptable to various working environments.

[0029] Additionally, in the pressure control assembly, the sealed box 28 is fixedly installed on one side of the camera device 27. Inside the sealed box 28, the top of the moving plate 29 is fixedly connected to two air storage bags 30. The top of the air storage bags 30 is fixedly connected to the top of the inner cavity of the sealed box 28. A hose 31 is provided on the top of the air storage bags 30. The hose 31 is fixedly installed on the top of the sealed box 28 and communicates with the interior of the air storage bag 30 on the same side. One end of the hose 31 is fixedly connected to the adjacent external threaded component 22, and the air guide groove 32 provided on one side of the hose 31 is opened on the rear side of the external threaded component 22. The air storage bag 30 communicates with the interior of the external threaded component 22 on the same side through the hose 31.

[0030] Furthermore, the lead screw 34, which rotates through the moving plate 29, is threadedly connected to the moving plate 29. The top of the lead screw 34 extends to the top of the sealing box 28 and is fixedly connected to a cross button 38. Two guide vertical rods 33, which pass through the moving plate 29, are fixedly installed inside the sealing box 28. Under the action of the two guide vertical rods 33, the moving plate 29 moves up and down inside the sealing box 28, pushing the rubber pad 23, which is fixedly connected to one side of the external threaded part 22, to press against the clean side wall of the vehicle. Because the hexagonal rod 25 is slidably inserted into the mounting bracket 26, the two rubber pads 23 can not be on the same plane. The two deformable rubber pads 23 can simultaneously press against the outer wall of the vehicle, applying a certain thrust to the camera device 27 to deform the two rubber pads 23 and seal the inside of the external threaded part 22. Rotating the cross knob 38 clockwise causes the lead screw 34 to rotate clockwise, driving the moving plate 29 to move horizontally downwards. The downward movement of the moving plate 29 causes the air storage bag 30 to unfold and draw air from the inside of the external threaded component 22 through the hose 31. The inside of the external threaded component 22 enters a negative pressure state. Multiple filling plates 24 fixedly connected inside the external threaded component 22 reduce the amount of air stored inside the external threaded component 22, increasing the speed at which the negative pressure inside the external threaded component 22 rises. As the moving plate 29 moves downwards, the increased negative pressure inside the external threaded component 22 adheres and fixes it to the outer wall of the vehicle. Finally, the camera device 27 is installed and suspended on the outside of the vehicle. This allows for quick and convenient replacement of the vehicle 1 equipped with the camera device 27, adapting to road data collection work in different environments and increasing the flexibility of using the detection equipment mounted on the side of the vehicle.

[0031] Two friction components 35 are fixedly installed on the outside of the lead screw 34. Friction components 36 and 37 are respectively sleeved on the outside of the two friction components 35. Friction components 36 and 37 are fixedly installed inside the sealed box 28. There is a certain friction between friction components 36 and 37 and friction components 35, which restricts the rotation of the lead screw 34 and prevents the lead screw 34 from rotating due to accidental vibration, thus ensuring the stability of the external thread component 22 installed on the side wall of the vehicle.

[0032] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for detecting unstructured road edges, characterized in that: Step 1: The original image is fed into the ResNet50 backbone network, and convolution and downsampling are performed layer by layer to output features at different levels. Step 2: Input the deep features output by the backbone network into the Deformable.ASPP module to output multi-scale augmented features with a fixed number of channels; Step 3: Feed the high-level features of ASPP into the edge enhancement module to extract edge features, and then stitch and fuse the edge features with the ASPP output in the channel dimension. Step 4: The edge-enhanced features are input together with features from different levels into a multi-level fusion decoder to unify the features from different levels to the same channel dimension and then splice them together. Step 5: Feed the decoded and fused features into the classifier, sample them to the same resolution as the input image to form a class response map, and then use bilinear interpolation.

2. The unstructured road edge detection method according to claim 1, characterized in that: Step one: After the acquired original image enters the ResNet50 backbone network, during the layer-by-layer convolution and downsampling process, the network outputs features at different levels: shallow features (from layer 1) retain detailed information such as edges and textures; mid-level features (from layers 2 and 3) take into account certain semantic and structural information; and deep features (from layer 4) contain global context and high-level semantics.

3. The unstructured road edge detection method according to claim 1, characterized in that: Step two: After obtaining the deep features output by the backbone network, they are input into the Deformable.ASPP module. This module consists of a 1×1 convolutional branch, multiple deformable convolutional branches, and a global pooling branch. The deformable convolutional branches can adaptively capture irregular shapes and complex boundaries by learning the sampling position offset. The global pooling branch introduces global context information. At the same time, the module is designed with a dynamic weight controller to adaptively adjust the response intensity of each branch according to the input features. Finally, all branch features are concatenated in the channel dimension and fused by convolution to output multi-scale enhanced features with a fixed number of channels. Among them, ASPP runs a dynamic hole rate mechanism. To address the problem of significant differences in the scale of unstructured roads, it proposes to introduce a dynamic hole rate into the ASPP module. The dilation coefficient of the dilated convolution is adaptively adjusted according to the content of the input feature map, avoiding the loss of semantic information caused by a fixed hole rate, thereby enhancing the model's ability to perceive road structures at different scales. The ASPP running branch weighting strategy suppresses redundant features and highlights key scale information, enabling the model to have a stronger ability to express multi-scale features in complex backgrounds.

4. The unstructured road edge detection method according to claim 1, characterized in that: Step 3: The high-level features obtained from ASPP are fed into the edge enhancement module. The edge enhancement module first uses a 3×3 convolution (similar to Sobel convolution) to extract potential edge features, and then concatenates the edge features with the ASPP output in the channel dimension. The concatenated features are then fused through an additional convolution block, thereby enhancing semantic information while highlighting the feature expression of the edge region. The edge enhancement module strengthens the representation of road boundary features, alleviates the problems of boundary blurring and breakage in unstructured road segmentation, and significantly improves the accuracy and continuity of boundary detection.

5. The unstructured road edge detection method according to claim 1, characterized in that: Step four: The edge-enhanced features are input into the multi-level fusion decoder along with the features from layer 1, layer 2 and layer 3. The decoder first unifies the features from different layers to the same channel dimension through a series of 1×1 convolutions. Then, it upsamples them to the same spatial resolution as layer 1 using bilinear interpolation. Next, all features are concatenated in the channel dimension and then fused through two layers of 3×3 convolutions to achieve complementarity between shallow details and deep semantics. The multi-level fusion decoder described in this invention differs from the single high-low layer feature fusion of the traditional DeepLabv3+. The multi-level feature fusion decoding structure designed in this invention ensures the consistency of the feature map before fusion through projection layers and upsampling, and fuses deep semantic information with shallow fine-grained features, which not only maintains global consistency, but also enhances the ability to restore details of local boundaries and textures.

6. The method for detecting unstructured road edges according to claim 1, characterized in that: Step 5: The decoded and fused features are fed into the classifier, and a class response map is generated using 1×1 convolution. The number of channels is equal to the number of classes in the segmentation task. Finally, bilinear interpolation is used to upsample the result to the same resolution as the input image to obtain the segmentation prediction map.

7. A detection device mounted on the side of a vehicle, applicable to the unstructured road edge detection method as described in any one of claims 1-6, comprising a vehicle (1) providing motion capability and a camera device (27) for capturing original images, characterized in that: The vehicle (1) is equipped with a support assembly on its top. The support assembly drives two internal threaded parts (15). Both internal threaded parts (15) are internally threaded with external threaded parts (22). A rubber pad (23) is fixedly connected to one side of the external threaded part (22). A hexagonal rod (25) and a mounting bracket (26) are provided between the other side of the external threaded part (22) and the camera device (27). A pressure control assembly is provided between the two external threaded parts (22).

8. The detection device mounted on the side of a vehicle according to claim 7 is characterized in that: The support assembly includes a robotic arm (2) fixedly mounted on the top of the vehicle (1), a pneumatic cylinder (4) fixedly connected to the top of the robotic arm (2), a mounting frame one (5) fixedly connected to the piston end of the pneumatic cylinder (4), a forward and reverse motor (6) fixedly connected to one side of the mounting frame one (5), a drive shaft (7) fixedly connected to the output end of the forward and reverse motor (6), a mounting frame three (10) fixedly connected to one end of the drive shaft (7), a mounting frame four (13) fixedly connected to the mounting frame three (10), and a mounting frame five (21) fixedly connected to one side inside the mounting frame four (13). Two telescopic rods (3) are fixedly connected to the top of the robotic arm (2). The piston end of the telescopic rod (3) is fixedly connected to the mounting frame one (5). A flat needle roller bearing (11) is fixedly connected to one side of the mounting frame one (5). Two connecting rods (12) are fixedly connected between the flat needle roller bearing (11) and the mounting frame three (10). A brake is sleeved on the outside of the drive shaft (7). 9), the brake (9) and the mounting bracket 1 (5) are fixedly connected by two mounting brackets 2 (8), the mounting bracket 5 (21) and the mounting bracket 4 (13) are rotatably connected by an operating rod (39) and two support shafts 2 (18), the operating rod (39) is fixedly installed with a gear 3 (20) on the outside, the two support shafts 2 (18) are fixedly installed with gears 2 (19) on the outside, the two gears 2 (19) are meshed with gears 3 (20), a friction element 1 (17) is provided on one side of the gear 2 (19), the friction element 1 (17) is fixedly connected to the mounting bracket 4 (13), the two internal threaded parts (15) are rotatably passed through the mounting bracket 5 (21), the internal threaded parts (15) are fixedly connected with a gear 1 (16) on the outside, the gear 1 (16) is meshed with the adjacent gear 2 (19), the internal threaded parts (15) and the mounting bracket 4 (13) are rotatably connected by a support shaft 1 (14).

9. The detection device mounted on the side of a vehicle according to claim 7 is characterized in that: Multiple filler plates (24) are fixedly connected inside the external threaded part (22). The hexagonal rod (25) is slidably inserted inside the mounting frame six (26). One end of the hexagonal rod (25) is fixedly connected to the external threaded part (22). The mounting frame six (26) is fixedly connected to the camera equipment (27).

10. The detection device mounted on the side of a vehicle according to claim 7 is characterized in that: The pressure control assembly includes a sealed box (28) fixedly connected to one side of the camera device (27), a moving plate (29) set inside the sealed box (28), two air-storing bladders (30) fixedly connected to the top of the moving plate (29), and a lead screw (34) rotatably passing through the moving plate (29). The lead screw (34) is threadedly connected to the moving plate (29), and the top end of the lead screw (34) extends to the top of the sealed box (28). A cross button (38) is fixedly connected to the top end of the lead screw (34). Two friction elements (35) are fixedly installed on the outside of the lead screw (34). Friction elements (36) and (37) are respectively sleeved on the outside of the two friction elements (35). Friction component three (36) and friction component four (37) are both fixedly installed inside the sealed box (28). Two guide rods (33) are provided on the moving plate (29). The guide rods (33) are fixedly installed inside the sealed box (28). The top of the air storage bag (30) is fixedly connected to the top of the inner cavity of the sealed box (28). A hose (31) is provided on the top of the air storage bag (30). The hose (31) is fixedly installed on the top of the sealed box (28). One end of the hose (31) is fixedly connected to the adjacent external threaded component (22). An air guide groove (32) is provided on one side of the hose (31). The air guide groove (32) is opened on the rear side of the external threaded component (22).

Images