Underwater embedded object recovery and classification system based on visual-tactile multi-feature perception

Abstract

The present invention discloses an underwater embedded object recovery and classification system based on visual-tactile multi-feature perception, including a flexible mechanical claw, a classification cabin, a main control computer, and an underwater camera. The system can be installed on an underwater embedded object collection robot to assist in the identification, recovery, and classification of underwater embedded objects. Through the phenomenon that human "visual deviation depends on tactile information correction", a visual-tactile multi-feature perception recognition method is proposed and applied to underwater embedded object recovery and classification projects. The shortcomings of the traditional visual recognition method, which has poor recognition accuracy under harsh working conditions and high requirements for the visual system, are improved. Through the multi-level recognition process of visual recognition, tactile recognition, and comprehensive recognition, the recognition accuracy is improved, and the requirements of each level of recognition method on equipment and environment are reduced. The system can be applied to various harsh working conditions of underwater embedded object recovery and classification projects.

Description

Technical Field

[0001] The present invention belongs to the field of underwater robot intelligent identification, and in particular relates to an underwater embedded object recovery and classification system based on visual-tactile multi-feature perception. Background Art

[0002] With the development of society and the exploration of the ocean, the demand for intelligent grasping of underwater garbage, seabed minerals, underwater aquatic products and other embedded objects is increasing. The development of related technologies has great significance for the environment, economy and people's livelihood. However, the current recognition system for grasping underwater embedded objects relies on visual recognition, which is difficult to meet the harsh underwater working conditions. Summary of the invention

[0003] Purpose of the invention: In order to solve the above problems, tactile information is used to correct visual deviations, and a visual-tactile multi-feature perception recognition system and method are proposed, which are applied to the underwater embedded object recovery and classification engineering.

[0004] Technical solution: The underwater embedded object recovery and classification system based on visual-tactile multi-feature perception consists of a flexible mechanical claw, a classification cabin, a main control computer, and an underwater camera. It can be installed on an underwater embedded object collection robot to assist in the identification, recovery, and classification of underwater embedded objects. The flexible mechanical claw, underwater camera, and main control computer are connected by cables and are respectively connected to the power supply system of the underwater embedded object collection robot. The flexible mechanical claw can place different types of embedded objects into the corresponding classification cabin.

[0005] Furthermore, the flexible mechanical claw is composed of a control arm, a flexible clamp, and a tactile sensor array, wherein the control arm is composed of a control head, a control keel, and a fastening bolt. The control head is used to transmit control signals and tactile characteristic signals, and to control the control keel to perform grasping and loosening work. The fastening bolt is used to fasten the flexible clamp, and the flexible clamp is nested on the control keel and is connected by a secondary reinforcement bolt. The tactile sensor array is adhered to the inner layer of the waterproof and corrosion-resistant flexible film on the inner side of the flexible clamp, and includes hardness sensors, pressure sensors, texture sensors, and optical fibers. The hardness sensors, pressure sensors, and texture sensors are alternately arrayed and connected by optical fibers. The hardness sensor, pressure sensor and texture sensor are basically composed of an upper electrode layer, a tactile chip layer, a lower electrode layer and an adhesive layer. The upper electrode layer and the lower electrode layer are flexible electrodes for transmitting the electrical signals generated by the contact of the tactile chip layer with the embedded material. The adhesive layer is used to provide an adhesive plane for the upper electrode layer, the tactile chip layer and the lower electrode layer, and can be adhered to the inner layer of the waterproof and corrosion-resistant flexible film on the inner side of the flexible clip. The tactile chip layer corresponds to the hardness chip layer, the pressure chip layer and the texture chip layer according to different information acquisition requirements.

[0006] Furthermore, the transmission system is composed of a tactile sensor array, a USB array, a USB integrator, a main control computer and several optical fibers. The tactile sensor array is used to obtain the tactile characteristic information of each contact point of the embedded object, which is transmitted by the optical fiber through the USB array. The USB integrator collects the tactile characteristic information of each contact point of the embedded object and transmits it to the main control computer through the optical fiber to sort out and obtain the tactile characteristic information of each contact point of the embedded object to avoid information mixing.

[0007] A method for recovering and classifying underwater embedded objects based on visual-tactile multi-feature perception includes the following steps:

[0008] A1. The underwater camera captures the target embedded object and transmits the image information to the main control computer;

[0009] A2. The main control computer pre-processes the target embedding image;

[0010] A3. The main control computer extracts image features;

[0011] A4. The main control computer performs image feature matching and determines the confidence level;

[0012] A5. The main control computer determines whether the image confidence level A is within the confidence interval;

[0013] A5.1 When the confidence level A of the single-chip judgment image is not within the confidence interval, the next target embedded object grasping workflow is entered, and the cycle starts from step A1;

[0014] A5.2 When the main control computer determines that the image confidence level A is within the confidence interval, the process proceeds to step A6;

[0015] A6. The main control computer controls the flexible mechanical claw to grab the target embedded object;

[0016] A7. The pressure sensor continuously transmits pressure information to the main control computer;

[0017] A8. The main control computer determines whether the pressure value is greater than or equal to the contact pressure;

[0018] A8.1 When the main control computer determines that the pressure value is less than the contact pressure, the cycle starts from step A6;

[0019] A8.2 When the main control computer determines that the pressure value is greater than or equal to the contact pressure, it proceeds to step A9;

[0020] A9. The tactile sensor array obtains the hardness and texture information of the target embedded object and transmits it to the main control computer;

[0021] A10. The main control computer extracts the hardness and texture information characteristics of the target embedded object;

[0022] A11. The main control computer performs target hardness and texture information feature matching;

[0023] A12. The main control computer determines whether the tactile confidence level B is within the confidence interval;

[0024] A12.1 When the main control computer determines that the tactile confidence B is not within the confidence interval, the captured embedded object is discarded and the next target embedded object capture workflow is entered, and the cycle starts from step A1;

[0025] A12.2 When the main control computer determines that the tactile confidence level B is within the confidence interval, the process proceeds to step A13;

[0026] A13. The main control computer calculates the comprehensive confidence C = a·A+b·B (a, b are weight coefficients, a+b=1);

[0027] A14. The main control computer determines whether the comprehensive confidence level C is within the confidence interval;

[0028] A14.1 When the main control computer determines that the comprehensive confidence C is not within the confidence interval, the captured embedded object is discarded and the next target embedded object capture workflow is entered, and the cycle starts from step A1;

[0029] A14.2 When the main control computer determines that the comprehensive confidence level C is within the confidence interval, the process proceeds to step A15;

[0030] A15. The main control computer determines the type of the embedded object and controls the manipulator to send the target embedded object into the corresponding classification cabin;

[0031] A16 Whether the main control computer receives the instruction to stop the recovery of embedded materials;

[0032] A16.1 When the main control computer does not receive the instruction to stop the embedding recovery work, it enters the next target embedding grabbing work flow and starts the cycle from step A1;

[0033] A16.2 When the main control computer receives the instruction to stop the recovery of embedded materials, it enters step A17;

[0034] A17. The main control computer controls the underwater embedded object grabbing robot to return.

[0035] Furthermore, the A2. main control computer pre-processes the target embedding image. In order to improve the image quality, reduce noise and enhance the target features, the color shift is first corrected by adjusting the white balance of the image to make the image more realistic. Then, the contrast of the image is enhanced and the noise in the image is removed by using techniques such as histogram equalization and filters to make the target more visible. Finally, according to the characteristics of the underwater environment, the Laplace enhancement algorithm is used to enhance the image detail features.

[0036] Furthermore, the main control computer extracts image features in A3. Step A1 is affected by the underwater environment. The scattering and absorption of underwater light poses a huge challenge to feature extraction. In order to improve the accuracy of underwater image target detection and recognition, the SURF algorithm can detect key points and corresponding local descriptors, so that the system has the advantages of fast speed and good noise resistance.

[0037] Furthermore, the A4. main control computer performs image feature matching to determine the confidence level. A convolutional neural network model (CNN) is loaded into the system, and the extracted features are input into the neural network model. The similarity between features is learned by the model, and a traditional nearest neighbor search matching method is used to calculate the matching result based on the feature similarity to determine the image confidence level.

[0038] Furthermore, the A8. main control computer determines whether the pressure value is greater than or equal to the contact pressure. The pressure sensor on the flexible mechanical claw is brought into contact with the embedded object, and the contact pressure data is collected in real time. The pressure data is converted from analog signals to digital signals and input into the single chip microcomputer, and the average pressure value within a specific time period is extracted, and then compared with the standard contact pressure measured in the experiment. When the pressure value is greater than the contact pressure, it indicates that the flexible mechanical claw is in successful contact with the embedded object, and the measured tactile data is true and reasonable.

[0039] Furthermore, the A13. main control computer calculates the comprehensive confidence C=a·A+b·B (a, b are weight coefficients, a+b=1). Data is collected and normalized for each tactile sensor on the mechanical claw, and the data is mapped to the same range to ensure that the data of different sensors are comparable. A weight is assigned to each sensor according to the importance and reliability of the sensor. The weight can be determined according to different working conditions through expert experience, experimental analysis or other feasible methods. More important and reliable sensors should be assigned higher weights, while more uncertain or unreliable sensors can be assigned lower weights. Finally, it is multiplied by the corresponding weight to obtain the comprehensive confidence.

[0040] Furthermore, the main control computer A15 determines the type of the embedded object and controls the manipulator to send the target embedded object into the corresponding classification compartment. A14 performs a comprehensive confidence judgment and determines the type of the embedded object by the comprehensive confidence value, so as to place the embedded object into the corresponding classification compartment.

[0041] Beneficial effects: This patent realizes the identification and classification of embedded objects (ore, garbage, aquatic products, etc.) under harsh underwater working conditions, thereby improving work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the system composition of the present invention;

[0043] Figure 2 This is a schematic diagram of the structure of the flexible mechanical claw of the present invention;

[0044] Figure 3 This is a schematic diagram of the flexible mechanical claw skeleton composition of the present invention;

[0045] Figure 4 It is a schematic diagram of the internal arrangement of the flexible mechanical claw of the present invention;

[0046] Figure 5 It is a schematic diagram of the structure of the tactile sensor of the present invention;

[0047] Figure 6 It is a schematic diagram of the composition of the multi-information integration system of the present invention;

[0048] Figure 7 It is a flow chart of the system operation of the present invention;

[0049] Figure 8 This is a system information interaction diagram of the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0051] like Figure 1 As shown, the underwater embedded object recovery and classification system based on visual-tactile multi-feature perception is composed of a flexible mechanical claw 1, a classification cabin 2, a main control computer 3, and an underwater camera 4, which can be installed on an underwater embedded object collection robot to assist in the identification, recovery, and classification of underwater embedded objects. The flexible mechanical claw 1, the underwater camera 4, and the main control computer 3 are connected through cables, and are respectively connected to the power supply system of the underwater embedded object collection robot. The flexible mechanical claw 1 can put different types of embedded objects into the corresponding classification cabin 2.

[0052] like Figure 2 , Figure 3 , Figure 4 , Figure 5As shown, the flexible mechanical claw 1 is composed of a control arm 11, a flexible clamp 12, and a tactile sensor array 13, wherein the control arm 11 is composed of a control head 111, a control keel 112, and a fastening bolt 113. The control head 111 is used to transmit control signals and tactile characteristic signals, and control the control keel 112 to perform grasping and loosening work, and the fastening bolt 113 is used to fasten the flexible clamp 12, and the flexible clamp 12 is nested on the control keel 112 and is connected by the fastening bolt 113 for secondary reinforcement. The tactile sensor array 13 is adhered to the inner layer of the waterproof and corrosion-resistant flexible film 121 on the inner side of the flexible clamp 12, and includes a hardness sensor 131, a pressure sensor 132, a texture sensor 133, and an optical fiber 134. The hardness sensor 131, the pressure sensor 132, and the texture sensor 133 are alternately arrayed and connected by the optical fiber 134. The hardness sensor 131, the pressure sensor 132, and the texture sensor 133 are basically composed of an upper electrode layer 231, a tactile chip layer 232, a lower electrode layer 233, and an adhesive layer 234. The upper electrode layer 231 and the lower electrode layer 233 are flexible electrodes for transmitting the electrical signal generated by the tactile chip layer 232 contacting the embedded object. The adhesive layer 234 is used to provide an adhesive plane for the upper electrode layer 231, the tactile chip layer 232, and the lower electrode layer 233, and can be adhered to the inner layer of the waterproof and corrosion-resistant flexible film 121 on the inner side of the flexible clip 12. The tactile chip layer 232 corresponds to the hardness chip layer, the pressure chip layer, and the texture chip layer according to different information acquisition requirements.

[0053] like Figure 6 As shown, the transmission system is composed of a tactile sensor array 13, a USB array 14, a USB integrator 15, a main control computer 3 and a number of optical fibers. The tactile sensor array 13 is used to obtain the tactile characteristic information of each contact point of the inlay, which is transmitted by the optical fiber through the USB array 14, and the USB integrator 15 collects the tactile characteristic information of each contact point of the inlay and transmits it to the main control computer 3 through the optical fiber to sort out and obtain the tactile characteristic information of each contact point of the inlay to avoid information mixing.

[0054] like Figure 7 As shown, the workflow of the underwater embedded object recovery and classification system based on visual-tactile multi-feature perception includes:

[0055] A1. The underwater camera captures the target embedded object and transmits the image information to the main control computer;

[0056] A2. The main control computer pre-processes the target embedding image;

[0057] A3. The main control computer extracts image features;

[0058] A4. The main control computer performs image feature matching and determines the confidence level;

[0059] A5. The main control computer determines whether the image confidence level A is within the confidence interval;

[0060] A5.1 When the confidence level A of the single-chip judgment image is not within the confidence interval, the next target embedded object grasping workflow is entered, and the cycle starts from process 1;

[0061] A5.2 When the main control computer determines that the image confidence level A is within the confidence interval, it proceeds to process 6;

[0062] A6. The main control computer controls the flexible mechanical claw to grab the target embedded object;

[0063] A7. The pressure sensor continuously transmits pressure information to the main control computer;

[0064] A8. The main control computer determines whether the pressure value is greater than or equal to the contact pressure;

[0065] A8.1 When the main control computer determines that the pressure value is less than the contact pressure, the cycle starts from process 6;

[0066] A8.2 When the main control computer determines that the pressure value is greater than or equal to the contact pressure, it enters process 9;

[0067] A9. The tactile sensor array obtains the hardness and texture information of the target embedded object and transmits it to the main control computer;

[0068] A10. The main control computer extracts the hardness and texture information characteristics of the target embedded object;

[0069] A11. The main control computer performs target hardness and texture information feature matching;

[0070] A12. The main control computer determines whether the tactile confidence level B is within the confidence interval;

[0071] A12.1 When the main control computer determines that the tactile confidence B is not within the confidence interval, the captured embedded object is discarded and the next target embedded object capture workflow is entered, and the cycle starts from process 1;

[0072] A12.2 When the main control computer determines that the tactile confidence level B is within the confidence interval, the process proceeds to step 13;

[0073] A13. The main control computer calculates the comprehensive confidence C = a·A+b·B (a, b are weight coefficients, a+b=1);

[0074] A14. The main control computer determines whether the comprehensive confidence level C is within the confidence interval;

[0075] A14.1 When the main control computer determines that the comprehensive confidence C is not within the confidence interval, the captured embedded object is discarded and the next target embedded object capture workflow is entered, and the cycle starts from process 1;

[0076] A14.2 When the main control computer determines that the comprehensive confidence level C is within the confidence interval, it enters process 15;

[0077] A15. The main control computer determines the type of the embedded object and controls the manipulator to send the target embedded object into the corresponding classification cabin;

[0078] A16 Whether the main control computer receives the instruction to stop the recovery of embedded materials;

[0079] A16.1 When the main control computer does not receive the instruction to stop the embedded object recovery work, it enters the next target embedded object grabbing work process and starts the cycle from process 1;

[0080] A16.2 When the main control computer receives the instruction to stop the recovery of embedded materials, it enters process 18;

[0081] A17. The main control computer controls the underwater embedded object grabbing robot to return.

[0082] like Figure 7 , Figure 8 As shown, in the control process, A2. the main control computer pre-processes the target embedding image. In order to improve the image quality, reduce noise and enhance the target features, the color shift is first corrected by adjusting the white balance of the image to make the image more realistic. Then, the contrast of the image is enhanced and the noise in the image is removed by using techniques such as histogram equalization and filters, making the target more visible. Finally, according to the characteristics of the underwater environment, the Laplace enhancement algorithm is used to enhance the image detail features.

[0083] The main control computer extracts image features in A3. Step A1 is affected by the underwater environment. The scattering and absorption of underwater light poses a huge challenge to feature extraction. In order to improve the accuracy of underwater image target detection and recognition, the SURF algorithm can detect key points and corresponding local descriptors, so that the system can obtain the advantages of fast speed and good noise resistance.

[0084] The A4. main control computer performs image feature matching and determines the confidence. A convolutional neural network model (CNN) is loaded into the system, and the extracted features are input into the neural network model. The similarity between features is learned by the model, and a traditional nearest neighbor search matching method is used to calculate the matching result based on the feature similarity to determine the image confidence.

[0085] The A8. main control computer determines whether the pressure value is greater than or equal to the contact pressure. The pressure sensor on the flexible mechanical claw is brought into contact with the embedded object, and the contact pressure data is collected in real time. The pressure data is converted from analog signals to digital signals and input into the single chip microcomputer, and the average pressure value within a specific time period is extracted, and then compared with the standard contact pressure measured in the experiment. When the pressure value is greater than the contact pressure, it indicates that the flexible mechanical claw is in successful contact with the embedded object, and the measured tactile data is true and reasonable.

[0086] The A13. main control computer calculates the comprehensive confidence C=a·A+b·B (a, b are weight coefficients, a+b=1). Data is collected and normalized for each tactile sensor on the mechanical claw, and the data is mapped to the same range to ensure that the data of different sensors are comparable. A weight is assigned to each sensor according to the importance and reliability of the sensor. The weight can be determined according to different working conditions through expert experience, experimental analysis or other feasible methods. More important and reliable sensors should be assigned higher weights, while more uncertain or unreliable sensors can be assigned lower weights. Finally, it is multiplied by the corresponding weight to obtain the comprehensive confidence.

[0087] The main control computer A15 determines the type of the embedded object and controls the manipulator to send the target embedded object into the corresponding classification compartment. A14 performs comprehensive confidence judgment and determines the type of the embedded object by the comprehensive confidence value, so as to place the embedded object into the corresponding classification compartment.

Claims

1. An underwater embedded object recovery and classification system based on visual-tactile multi-feature perception, characterized in that: The invention comprises a flexible mechanical claw (1), a classification chamber (2), a main control computer (3), and an underwater camera (4), wherein the flexible mechanical claw (1), the underwater camera (4) and the main control computer (3) are connected via cables and are respectively connected to a power supply system of an underwater embedded object collection robot, and the flexible mechanical claw (1) can place different types of embedded objects into corresponding classification chambers (2); The underwater encrustation recovery and classification method using the underwater encrustation recovery and classification system based on visual-tactile multi-feature perception comprises the following steps: A1. The underwater camera captures the target embedded object and transmits the image information to the main control computer; A2. The main control computer pre-processes the target embedding image; A3. The main control computer extracts image features; A4. The main control computer performs image feature matching and determines the confidence level; A5. The main control computer determines whether the image confidence level A is within the confidence interval; A5.1 When the confidence level A of the single-chip judgment image is not within the confidence interval, the next target embedded object grasping workflow is entered, and the cycle starts from step A1; A5.2 When the main control computer determines that the image confidence level A is within the confidence interval, the process proceeds to step A6; A6. The main control computer controls the flexible mechanical claw to grab the target embedded object; A7. The pressure sensor continuously transmits pressure information to the main control computer; A8. The main control computer determines whether the pressure value is greater than or equal to the contact pressure; A8.1 When the main control computer determines that the pressure value is less than the contact pressure, the cycle starts from step A6; A8.2 When the main control computer determines that the pressure value is greater than or equal to the contact pressure, it proceeds to step A9; A9. The tactile sensor array obtains the hardness and texture information of the target embedded object and transmits it to the main control computer; A10. The main control computer extracts the hardness and texture information characteristics of the target embedded object; A11. The main control computer performs target hardness and texture information feature matching; A12. The main control computer determines whether the tactile confidence level B is within the confidence interval; A12.1 When the main control computer determines that the tactile confidence B is not within the confidence interval, the captured embedded object is discarded and the next target embedded object capture workflow is entered, and the cycle starts from step A1; A12.2 When the main control computer determines that the tactile confidence level B is within the confidence interval, the process proceeds to step A13; A13. The main control computer calculates the comprehensive confidence C = a·A+b·B, where a and b are weight coefficients, and a+b=1; A14. The main control computer determines whether the comprehensive confidence level C is within the confidence interval; A14.1 When the main control computer determines that the comprehensive confidence C is not within the confidence interval, the captured embedded object is discarded and the next target embedded object capture workflow is entered, and the cycle starts from step A1; A14.2 When the main control computer determines that the comprehensive confidence level C is within the confidence interval, the process proceeds to step A15; A15. The main control computer determines the type of the embedded object and controls the manipulator to send the target embedded object into the corresponding classification cabin; A16 Whether the main control computer receives the instruction to stop the recovery of embedded materials; A16.1 When the main control computer does not receive the instruction to stop the embedding recovery work, it enters the next target embedding grabbing work flow and starts the cycle from step A1; A16.2 When the main control computer receives the instruction to stop the recovery of embedded materials, it enters step A17; A17. The main control computer controls the underwater embedded object grabbing robot to return.

2. The underwater embedded object recovery and classification system based on visual-tactile multi-feature perception according to claim 1 is characterized in that: The flexible mechanical claw (1) comprises a control arm (11), a flexible clamp (12), and a tactile sensor array (13), wherein the control arm (11) is composed of a control head (111), a control keel (112), and a fastening bolt (113); the control head (111) is used to transmit control signals and tactile characteristic signals, and control the control keel (112) to perform grasping and loosening operations; the fastening bolt (113) is used to fasten the flexible clamp (12); the flexible clamp (12) is nested on the control keel (112) and is connected by a secondary reinforcement through the fastening bolt (113).

3. The underwater embedded object recovery and classification system based on visual-tactile multi-feature perception according to claim 2 is characterized in that: The tactile sensor array (13) is adhered to the inner layer of the waterproof and corrosion-resistant flexible film (121) on the inner side of the flexible clip (12), and the tactile sensor array (13) includes a hardness sensor (131), a pressure sensor (132), a texture sensor (133), and an optical fiber (134). The hardness sensor (131), the pressure sensor (132), and the texture sensor (133) are arranged in an alternating array and are connected by the optical fiber (134).

4. The underwater embedded object recovery and classification system based on visual-tactile multi-feature perception according to claim 3 is characterized in that: The hardness sensor (131), the pressure sensor (132), and the texture sensor (133) basically comprise an upper electrode layer (231), a tactile chip layer (232), a lower electrode layer (233), and an adhesive layer (234), wherein the upper electrode layer (231) and the lower electrode layer (233) are flexible electrodes for transmitting electrical signals generated when the tactile chip layer (232) contacts an embedded object; the adhesive layer (234) is used to provide an adhesive plane for the upper electrode layer (231), the tactile chip layer (232), and the lower electrode layer (233), and can be adhered to the inner layer of the waterproof and corrosion-resistant flexible film (121) on the inner side of the flexible clip (12); and the tactile chip layer (232) is respectively set as a hardness chip layer, a pressure chip layer, and a texture chip layer according to different information acquisition requirements.

5. The underwater embedded object recovery and classification system based on visual-tactile multi-feature perception according to claim 1 is characterized in that: The underwater embedded object recovery and classification system based on visual-tactile multi-feature perception also includes a transmission system, which includes a tactile sensor array (13), a USB array (14), a USB integrator (15), a main control computer (3) and a number of optical fibers. The tactile sensor array (13) is used to obtain tactile feature information of each contact point of the embedded object, which is transmitted by the optical fiber through the USB array (14), and the USB integrator (15) collects the tactile feature information of each contact point of the embedded object and transmits it to the main control computer (3) through the optical fiber to sort out and obtain the tactile feature information of each contact point of the embedded object.

6. The underwater embedded object recovery and classification system based on visual-tactile multi-feature perception according to claim 1 is characterized in that: The preprocessing in A2 first corrects the color shift by adjusting the white balance of the image, then uses histogram equalization and filters to enhance the contrast of the image, and finally uses the Laplace enhancement algorithm to enhance the image detail features according to the characteristics of the underwater environment.

7. The underwater embedded object recovery and classification system based on visual-tactile multi-feature perception according to claim 1 is characterized in that: The A4 main control computer performs image feature matching, determines the confidence, loads the convolutional neural network model into the system, and inputs the extracted features into the neural network model. The similarity between features is learned through the model, and the nearest neighbor search matching method is used to obtain the matching result based on the feature similarity calculation to determine the image confidence.

8. The underwater embedded object recovery and classification system based on visual-tactile multi-feature perception according to claim 1 is characterized in that: The A8 main control computer determines whether the pressure value is greater than or equal to the contact pressure, brings the pressure sensor on the flexible mechanical claw into contact with the embedded object, collects contact pressure data in real time, converts the pressure data from analog signals into digital signals and inputs them into the single chip microcomputer, extracts the average pressure value within a specific time period, and then compares it with the standard contact pressure measured experimentally.

9. The underwater embedded object recovery and classification system based on visual-tactile multi-feature perception according to claim 1 is characterized in that: The A13 main control computer calculates the comprehensive confidence, collects and normalizes data for each tactile sensor on the mechanical claw, maps the data to the same range, assigns a weight to each sensor, and multiplies the confidence by the corresponding weight to obtain the comprehensive confidence.

Images