Rapid path planning method and device suitable for lightweight unmanned aerial vehicle

Abstract

The invention discloses a rapid path planning method and device suitable for a lightweight unmanned aerial vehicle, and the method specifically comprises the steps: deploying the rapid path planning device suitable for the lightweight unmanned aerial vehicle, and carrying out the initialization; the visual collection module collects and compresses visual information images of the environment and transmits the images to the machine learning module. The machine learning module firstly compresses a neural network model through depth separable convolution, reverse residual block, channel scaling and quantization processing, then measures the compressed model according to an unmanned aerial vehicle path planning business index, and finally learns and tests the model to generate an optimized path planning strategy; and the strategy deployment module issues the path planning strategy generated by the machine learning module to the unmanned aerial vehicle for path planning. The device comprises a visual acquisition module, a machine learning module and a strategy deployment module. According to the invention, the energy consumption and memory occupation of the lightweight unmanned aerial vehicle are reduced, the speed of path planning is improved, and the high accuracy of path planning is maintained at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle control, and in particular to a fast path planning method and device suitable for a lightweight unmanned aerial vehicle. Background Art

[0002] In recent years, unmanned aerial vehicles, unmanned vehicles, unmanned ships and other drones have been widely used in logistics, mapping and disaster relief. Lightweight drones have broad prospects in applications due to their low cost and high flexibility. However, the path planning of lightweight drones in complex environments relies on real-time data processing, which requires efficient computing power and low-energy operation mode to cope with problems such as limited memory and energy sensitivity.

[0003] In practical applications, path planning is the core task of drone missions, which directly affects their safety and efficiency. Traditional path planning algorithms have low operating efficiency on small devices and it is difficult to balance computing accuracy and energy consumption requirements, making it more difficult for lightweight drones to make quick decisions in complex environments. To meet the above challenges, the path planning system needs to ensure the accuracy and real-time performance of the planning while reducing computing resource usage and energy consumption.

[0004] Patent CN 118587617 A discloses a method for tracking ground targets using a lightweight UAV. It uses a lightweight network mobive-v2 for feature extraction, optimizes the target tracking network, reduces memory usage, reduces the amount of calculation, makes the model lighter, and ensures tracking accuracy and speed. It can accurately track ground targets, but this method cannot perform path planning when the UAV performs tasks. Therefore, it is urgent to study an efficient path planning method for lightweight UAVs to achieve fast and accurate path planning for UAVs under resource-constrained conditions. Summary of the invention

[0005] The object of the present invention is to provide a lightweight UAV path planning method and device with low energy consumption, small memory requirement, fast path planning speed and high accuracy.

[0006] The technical solution to achieve the purpose of the present invention is: a fast path planning method suitable for lightweight unmanned aerial vehicles, comprising the following steps:

[0007] Step 1: deploy a fast path planning device suitable for lightweight UAVs, including a visual acquisition module, a machine learning module, and a strategy deployment module, and initialize the device;

[0008] Step 2: The visual acquisition module collects and compresses the visual information images of the environment and transmits them to the machine learning module;

[0009] Step 3: The machine learning module first compresses the neural network model through deep separable convolution, inverse residual block, channel scaling and quantization processing, then measures the compressed neural network model according to the UAV path planning business indicators, and finally learns and tests the compressed neural network model to generate an optimized path planning strategy;

[0010] Step 4: The strategy deployment module sends the path planning strategy generated by the machine learning module to the drone for path planning.

[0011] A fast path planning device suitable for lightweight unmanned aerial vehicles, the device is used to implement the fast path planning method suitable for lightweight unmanned aerial vehicles, the device includes a visual acquisition module, a machine learning module and a strategy deployment module, and the functions of each module are as follows:

[0012] The visual acquisition module collects and compresses the visual information images of the environment and transmits them to the machine learning module;

[0013] The machine learning module first compresses the neural network model through deep separable convolution, inverse residual block, channel scaling and quantization processing, then measures the compressed neural network model according to the UAV path planning business indicators, and finally learns and tests the compressed neural network model to generate an optimized path planning strategy;

[0014] The strategy deployment module sends the path planning strategy generated by the machine learning module to the drone for path planning.

[0015] Compared with the prior art, the present invention has the following significant advantages: (1) by optimizing the compression technology of machine learning and the path planning strategy, the computing resource occupation and energy consumption are reduced; (2) the accuracy and real-time performance of path planning are guaranteed, so that the UAV can achieve fast and accurate path planning under resource-constrained conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a structural schematic diagram of a fast path planning device suitable for a lightweight unmanned aerial vehicle according to the present invention.

[0017] Figure 2 The figure is a flow chart of a fast path planning method suitable for a lightweight unmanned aerial vehicle according to the present invention. DETAILED DESCRIPTION

[0018] The present invention provides a fast path planning method applicable to a lightweight unmanned aerial vehicle, comprising the following steps:

[0019] Step 1: deploy a fast path planning device suitable for lightweight UAVs, including a visual acquisition module, a machine learning module, and a strategy deployment module, and initialize the device;

[0020] Step 2: The visual acquisition module collects and compresses the visual information images of the environment and transmits them to the machine learning module;

[0021] Step 3: The machine learning module first compresses the neural network model through deep separable convolution, inverse residual block, channel scaling and quantization processing, then measures the compressed neural network model according to the UAV path planning business indicators, and finally learns and tests the compressed neural network model to generate an optimized path planning strategy;

[0022] Step 4: The strategy deployment module sends the path planning strategy generated by the machine learning module to the drone for path planning.

[0023] As a specific example, the drone includes an unmanned aerial vehicle, an unmanned vehicle, an unmanned ship and an unmanned underwater vehicle.

[0024] As a specific example, in step 2, the visual acquisition module collects and compresses the visual information image of the environment and transmits it to the machine learning module, as follows:

[0025] Step 2.1, the visual acquisition module uses the camera or lidar equipment installed on the drone to obtain real-time visual information images of the environment, including obstacles, terrain and dynamic objects around the drone;

[0026] Step 2.2, the visual acquisition module compresses the visual information image, compresses the visual information image whose clarity is higher than the threshold, and directly uses the original image for the visual information image whose clarity is not higher than the threshold;

[0027] Step 2.3: The visual acquisition module transmits the processed visual information image to the machine learning module.

[0028] As a specific example, in step 2.2, the visual information image with a clarity higher than the threshold is compressed as follows:

[0029] The resolution of the original visual information image is reduced while retaining the necessary visual information. The compression process is achieved by reducing the resolution, grayscale processing or image compression algorithm, where:

[0030] The resolution reduction is to reduce the resolution of the image from the original size;

[0031] The grayscale processing is to convert the color image into a grayscale image, reducing the information of each pixel from three channels of red, green and blue to one grayscale channel;

[0032] The image compression algorithm uses a JPEG or PNG image compression algorithm to reduce the size of an image file.

[0033] As a specific example, step 3 is as follows:

[0034] Step 3.1, the machine learning module uses deep separable convolution, reverse residual block, channel scaling and quantization processing to decompose the standard convolutional neural network into a lightweight model to compress the neural network model;

[0035] Step 3.2, measure the compressed neural network model according to the UAV path planning business indicators;

[0036] Step 3.3: The machine learning module learns and tests the compressed neural network model based on real-time images, generates an optimized path planning strategy, and passes it to the strategy deployment module.

[0037] As a specific example, the machine learning module in step 3.1 uses depthwise separable convolution, inverse residual block, channel scaling and quantization to decompose the standard convolutional neural network into a lightweight model as follows:

[0038] Step 3.1.1, the machine learning module splits the standard convolution operation into depth-separable convolution and point-by-point convolution. Suppose the input feature map is ,in is the height, is the width, is the number of channels, Indicates size The set of real matrices of ;

[0039] Depthwise separable convolution for each channel , the convolution kernel size is :

[0040]

[0041] in, represents the convolution operation, For the The convolution kernel of channels; For the The characteristics of each channel; For the The depth-wise separable convolution result of channels;

[0042] Point-wise convolution uses The convolution kernel linearly combines the features of each channel:

[0043]

[0044] in, for The weight matrix of the convolution kernel; is the linear combination result of each channel feature;

[0045] Step 3.1.2, the reverse residual block is divided into the following three steps:

[0046] (1) Channel expansion: through Convolution expands the number of channels to :

[0047]

[0048] in is the expansion coefficient, used to increase the feature space; is the expanded feature map; Expand the matrix for the channel; Indicates size The set of real matrices of ;

[0049] (2) Depth-wise separable convolution: for the expanded feature map conduct Depthwise separable convolution:

[0050]

[0051] in, is the depth convolution matrix; Indicates size The set of real matrices of ; for The depth-wise separable convolution result of ;

[0052] (3) Channel contraction: through Convolution restores the number of channels to the number of output channels :

[0053]

[0054] in, Recover the matrix for the channel; Indicates size The set of real matrices of ;

[0055] Step 3.1.3: Channel scaling by setting the scaling factor To reduce the number of channels per layer, set the number of channels to , then the number of channels after scaling for:

[0056]

[0057] Step 3.1.4: Quantization converts the model's weights and activations from floating point numbers to fixed point numbers, setting the original weight matrix to a floating point matrix , then the quantitative formula is:

[0058]

[0059] in, is the quantization scale factor, is the converted fixed-point matrix.

[0060] As a specific example, the compressed neural network model is measured according to the drone path planning business indicators described in step 3.2, as follows:

[0061] Step 3.2.1: In terms of drone business optimization, smooth motion control is achieved and the speed of the drone is and angular velocity For smoothing, a first-order low-pass filter is used, and the formula is as follows:

[0062]

[0063]

[0064] in, and is the filter coefficient, and its value range is ; , Speed , angular velocity No. Second filtering result; , Speed , angular velocity No. Second filtering result; , Speed , angular velocity Results without filtering;

[0065] Step 3.2.2. Loss function It includes two parts: mean square error MSE and binary cross entropy BCE:

[0066]

[0067] Among them, MSE is used to measure the error of the model when predicting continuous variables; BCE is used to deal with the loss of classification tasks. is the weight used to adjust the two parts of the loss.

[0068] As a specific example, the strategy deployment module described in step 4 sends the path planning strategy generated by the machine learning module to the drone for path planning, as follows:

[0069] Step 4.1, the strategy deployment module receives the path planning strategy generated by the machine learning module and parses it into low-level control instructions that the drone can recognize;

[0070] Step 4.2, encoding the parsed control instructions into machine language or a set communication format;

[0071] Step 4.3: The strategy deployment module sends instructions to the control module of the drone in real time according to the preset cycle or trigger condition. After sending, the strategy deployment module waits for the feedback signal from the drone to confirm whether the instruction is successfully executed. If it is executed, go to step 4.5, otherwise go to step 4.4;

[0072] Step 4.4: If the drone shows in the feedback that it has failed to fully execute the command, the strategy deployment module adjusts and resends the correction command within the preset time;

[0073] Step 4.5: If the UAV shows in the feedback that the command has been fully executed, the strategy deployment module continuously monitors the UAV's status information, including position, speed, and orientation, and returns the real-time environmental status to the machine learning module through a closed-loop feedback mechanism to update the model or re-plan the path.

[0074] As a specific example, in step 4.2, the parsed control instructions are encoded into machine language or a specific communication format, including reduced instruction set, binary encoding compression, multi-instruction package transmission, and parameterized encoding design;

[0075] The simplified instruction set is a simplified instruction set established for the commonly used control commands of drones;

[0076] The binary code compression is to compress the instructions using binary code, mapping each control instruction to a unique binary code;

[0077] The multi-command package transmission is to package multiple commands together and transmit them to the drone, reducing the frequency of command transmission one by one;

[0078] The parameterized coding design adopts a parameterized coding mode to specify multiple pieces of information through one instruction.

[0079] The present invention also provides a fast path planning device suitable for lightweight unmanned aerial vehicles, which is used to implement the fast path planning method suitable for lightweight unmanned aerial vehicles. The device includes a visual acquisition module, a machine learning module and a strategy deployment module, and the functions of each module are as follows:

[0080] The visual acquisition module collects and compresses the visual information images of the environment and transmits them to the machine learning module;

[0081] The machine learning module first compresses the neural network model through deep separable convolution, inverse residual block, channel scaling and quantization processing, then measures the compressed neural network model according to the UAV path planning business indicators, and finally learns and tests the compressed neural network model to generate an optimized path planning strategy;

[0082] The strategy deployment module sends the path planning strategy generated by the machine learning module to the drone for path planning.

[0083] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0084] Example

[0085] like Figure 1 As shown, this embodiment provides a fast path planning device suitable for a lightweight UAV, including a visual acquisition module, a machine learning module and a strategy deployment module;

[0086] The visual acquisition module uses the inherent visual acquisition equipment of the drone to collect and compress visual information;

[0087] The machine learning module uses a neural network model to learn based on visual information, compresses the convolutional neural network through deep separable convolution, reverse residual block, channel scaling and quantization processing technology to reduce memory usage and energy consumption, uses collision rate and steering angle as evaluation indicators for path planning, conducts learning and testing, and finally generates an optimized path planning strategy;

[0088] The strategy deployment module receives planning decisions from the machine learning module and maps them to the action module of the drone in the form of machine code or machine language to achieve real-time decision-making based on visual information.

[0089] like Figure 2 As shown, this embodiment also provides a fast path planning method suitable for a lightweight UAV, comprising the following steps:

[0090] Step 1: deploy a fast path planning device suitable for lightweight UAVs and initialize the device;

[0091] As a specific example, the lightweight drones include but are not limited to unmanned aerial vehicles, including multi-rotor drones, fixed-wing drones, and vertical take-off and landing fixed-wing drones; unmanned vehicles, including self-driving cars, unmanned transport vehicles, and robot vehicles, which are unmanned devices for ground navigation; unmanned ships, including unmanned patrol ships, unmanned rescue ships, and unmanned cargo ships, which are used for surface missions; and unmanned submersibles, including underwater robots and unmanned submarines, which are used for underwater navigation and detection missions. The drones are equipped with visual acquisition equipment, such as cameras and laser radars, and have basic computing and path adjustment capabilities.

[0092] As a specific example, the visual acquisition module uses the inherent visual acquisition equipment of the drone to collect and compress visual information, and transmits it to the machine learning module for processing;

[0093] The machine learning module uses a neural network model to learn based on visual information, compresses the convolutional neural network through deep separable convolution, reverse residual block, channel scaling and quantization processing technology to reduce memory usage and energy consumption, uses collision rate and steering angle as evaluation indicators for path planning, conducts learning and testing, and finally generates an optimized path planning strategy;

[0094] The strategy deployment module receives planning decisions from the machine learning module and maps them to the action module of the drone in the form of machine code or machine language to achieve real-time decision-making based on visual information.

[0095] Step 2: The visual acquisition module collects and processes environmental information and passes it to the machine learning module, as follows:

[0096] Step 2.1: The visual acquisition module uses the inherent camera or lidar equipment on the drone to obtain real-time visual information of the environment, including obstacles, terrain, and dynamic objects around the drone, to ensure that path planning has sufficient environmental data support;

[0097] As a specific example, if the acquisition device acquires visual information with high definition, such as images captured by a high-definition camera, the visual acquisition module compresses it into a small-size image; if the acquisition device has low definition, such as in poor lighting or low device resolution, the small-size or grayscale image is directly used for subsequent processing;

[0098] Step 2.2, the visual acquisition module performs image compression and format conversion. According to the clarity of the acquired image, the corresponding compression operation is performed. For high-definition images, appropriate resolution reduction and lossy compression algorithms are used to compress the image into a small-size format; for low-definition images, the original small-size or grayscale image is directly used to reduce additional processing steps;

[0099] As a specific example, the image is compressed into a small size format using appropriate resolution reduction and lossy compression algorithms, as follows:

[0100] Reduce the resolution of the original image to make it take up less storage space and retain the necessary visual information. The compression process includes reducing the resolution, adjusting the grayscale, and using the image compression algorithm, where:

[0101] The resolution reduction is to reduce a high-definition image with a large resolution to a small-size image with a small resolution;

[0102] Grayscale processing is to perform grayscale processing on the image, convert the color image into a grayscale image, reduce the information of each pixel from three channels of red, green and blue to one grayscale channel, and further reduce the amount of data;

[0103] The image compression algorithm described above uses a JPEG or PNG image compression algorithm to reduce the file size, which is specifically divided into lossy compression and lossless compression. Lossy compression can significantly reduce the file size and is suitable for tasks that do not require fine details; lossless compression retains all information of the image, but has a lower compression rate.

[0104] Step 2.3: The visual acquisition module transmits the processed image information to the machine learning module.

[0105] As a specific example, the image information transmission method uses a frame mode with more than 25 frames per second to ensure the real-time performance of image processing.

[0106] Step 3: The machine learning module uses a lightweight model to learn and outputs a path planning strategy, as follows:

[0107] Step 3.1: The machine learning module uses depthwise separable convolution, inverse residual block, channel scaling, and quantization, which can also be used in combination to decompose the standard convolutional neural network into a lightweight model, as follows:

[0108] Step 3.1.1, the machine learning module greatly reduces the computational complexity by splitting the standard convolution operation into depth-separable convolution and point-by-point convolution. Suppose the input feature map is ,in is the height, is the width, is the number of channels;

[0109] Depthwise separable convolution for each channel , the convolution kernel size is :

[0110]

[0111] in, represents the convolution operation, For the The computational complexity of depth-wise separable convolution is approximately .

[0112] Point-wise convolution uses The convolution kernel linearly combines the features of each channel:

[0113]

[0114] in, for The weight matrix of the convolution kernel; the computational cost of point-by-point convolution is ,in is the number of output channels.

[0115] Compared with the standard convolution, the computational complexity is , the total computational cost of depthwise separable convolution is , the computational complexity of depth-wise separable convolution is significantly reduced.

[0116] Step 3.1.2, the reverse residual block is divided into the following three steps:

[0117] (1) Channel expansion: through Convolution expands the number of channels to :

[0118]

[0119] in is the expansion factor, which is used to increase the feature space in order to extract more information;

[0120] (2) Depth-wise separable convolution: Perform convolution on the expanded feature map Depthwise separable convolution:

[0121]

[0122] (3) Channel contraction: through Convolution restores the number of channels to the number of output channels :

[0123]

[0124] The total computational effort of the reverse residual block is , further reducing computing requirements, especially suitable for embedded systems.

[0125] Step 3.1.3: Channel scaling by setting the scaling factor To reduce the number of channels in each layer, thereby reducing the number of model parameters and calculations, the original number of channels is set to , then the number of channels after scaling is:

[0126]

[0127] All occurrences of The part will be replaced by ,choose , the memory requirement of the model is only 1 / 168 of the original.

[0128] Step 3.1.4: Quantization converts the model's weights and activations from floating point numbers to fixed point numbers, setting the original weight matrix to a floating point matrix , then the quantitative formula is:

[0129]

[0130] in, is the quantization scaling factor, which is determined by minimizing the quantization error. After quantization, the storage requirement of each weight is only 1 / 4 of the original. For shallower lightweight networks, bypass connections have limited performance improvement and can therefore be removed.

[0131] Step 3.2: The machine learning module is measured based on the drone path planning business indicators, as follows:

[0132] Step 3.2.1: In terms of drone business optimization, smooth motion control is achieved and the speed of the drone is and angular velocity For smoothing, a first-order low-pass filter is used, and the formula is as follows:

[0133]

[0134]

[0135] in, and is the number of filter coefficients, and its value range is , the larger A smaller value will make the filtering effect weaker but more responsive; A value of 0 will make the output smoother but the response slower;

[0136] Step 3.2.2, the loss function includes two parts: mean square error MSE and binary cross entropy BCE:

[0137]

[0138] Among them, MSE is used to measure the error of the model when predicting continuous variables; BCE is used to deal with the loss of classification tasks. Used to adjust the weights of the two parts of the loss.

[0139] Step 3.3: The machine learning module generates a path planning strategy based on the real-time image and passes it to the strategy deployment module.

[0140] Step 4: The policy deployment module sends the policy generated by the machine learning module to the drone for control, as follows:

[0141] Step 4.1, the strategy deployment module receives the path planning strategy generated by the machine learning module and parses it into low-level control instructions that the drone can recognize;

[0142] As a specific example, the path planning strategy involves heading adjustment, speed change, and obstacle avoidance operations. The deployment module needs to convert high-level strategies, such as "shift to the right to avoid obstacles", into specific control commands, such as "turn right at an angle of 10 degrees and maintain a distance of 2 meters";

[0143] Step 4.2: Encode the parsed control instructions into machine language or a specific communication format. For lightweight UAVs with limited memory, the instruction encoding should be as concise as possible, and possible instruction redundancy should be taken into account to optimize the transmission efficiency of the control signal, including reduced instruction set, binary encoding compression, multi-instruction package transmission, and parameterized encoding design;

[0144] The simplified instruction set is to establish a simplified instruction set for the commonly used control commands of the drone, such as simplifying "turn right 10 degrees" to "R10", or simplifying "maintain current speed" to "VS";

[0145] The binary code compression is to compress the instructions using binary code, and map each control instruction to a unique binary code, such as representing the direction, speed and distance information with different binary fields, similar to 0001 1010 1100 representing "turn right 10 degrees and move forward at a speed of 3 meters per second".

[0146] The multi-command package transmission is to package multiple commands together and transmit them to the drone, reducing the frequency of transmitting the commands one by one. For example, if the drone needs to execute "turn left 10 degrees, move forward 2 meters, turn right 5 degrees" in sequence on the path, the commands can be compressed into a package format, and multiple commands can be divided in the data packet using a marker bit or a separator;

[0147] The parameterized coding design adopts a parameterized coding mode to specify multiple pieces of information through one instruction, such as "P5, R10" means "turn right 10 degrees at the 5th key point in the path".

[0148] Step 4.3: The strategy deployment module sends instructions to the control module of the drone in real time according to the preset cycle or trigger condition. After sending, the strategy deployment module waits for the feedback signal from the drone to confirm whether the instruction is successfully executed. If it is executed, go to step 4.5, otherwise go to step 4.4;

[0149] Step 4.4: If the drone shows in the feedback that it has failed to fully execute the command, the strategy deployment module adjusts and resends the correction command within the preset time;

[0150] Step 4.5: If the UAV shows in the feedback that the command has been fully executed, the strategy deployment module continuously monitors the UAV's status information, including position, speed, and orientation, and returns the real-time environmental status to the machine learning module through a closed-loop feedback mechanism to update the model or re-plan the path.

[0151] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A fast path planning method suitable for lightweight unmanned aerial vehicles, characterized in that: The following steps are involved: Step 1: deploy a fast path planning device suitable for lightweight UAVs, including a visual acquisition module, a machine learning module, and a strategy deployment module, and initialize the device; Step 2: The visual acquisition module collects and compresses the visual information images of the environment and transmits them to the machine learning module; Step 3: The machine learning module first compresses the neural network model through deep separable convolution, inverse residual block, channel scaling and quantization processing, then measures the compressed neural network model according to the UAV path planning business indicators, and finally learns and tests the compressed neural network model to generate an optimized path planning strategy; Step 4: The strategy deployment module sends the path planning strategy generated by the machine learning module to the drone for path planning.

2. The fast path planning method for a lightweight unmanned aerial vehicle according to claim 1, characterized in that: The drones include unmanned aerial vehicles, unmanned vehicles, unmanned ships and unmanned submarines.

3. The fast path planning method for a lightweight UAV according to claim 1, characterized in that: In step 2, the visual acquisition module collects and compresses the visual information image of the environment and passes it to the machine learning module as follows: Step 2.1, the visual acquisition module uses the camera or lidar equipment installed on the drone to obtain real-time visual information images of the environment, including obstacles, terrain and dynamic objects around the drone; Step 2.2, the visual acquisition module compresses the visual information image, compresses the visual information image whose clarity is higher than the threshold, and directly uses the original image for the visual information image whose clarity is not higher than the threshold; Step 2.3: The visual acquisition module transmits the processed visual information image to the machine learning module.

4. The fast path planning method for a lightweight unmanned aerial vehicle according to claim 2, characterized in that: In step 2.2, the visual information image with a clarity higher than the threshold is compressed as follows: The resolution of the original visual information image is reduced while retaining the necessary visual information. The compression process is achieved by reducing the resolution, grayscale processing or image compression algorithm, where: The resolution reduction is to reduce the resolution of the image from the original size; The grayscale processing is to convert the color image into a grayscale image, reducing the information of each pixel from three channels of red, green and blue to one grayscale channel; The image compression algorithm uses a JPEG or PNG image compression algorithm to reduce the size of an image file.

5. The fast path planning method for a lightweight UAV according to claim 1, characterized in that: The step 3 is as follows: Step 3.1, the machine learning module uses deep separable convolution, reverse residual block, channel scaling and quantization processing to decompose the standard convolutional neural network into a lightweight model to compress the neural network model; Step 3.2, measure the compressed neural network model according to the UAV path planning business indicators; Step 3.3: The machine learning module learns and tests the compressed neural network model based on real-time images, generates an optimized path planning strategy, and passes it to the strategy deployment module.

6. The fast path planning method for a lightweight UAV according to claim 5, characterized in that: In step 3.1, the machine learning module uses depthwise separable convolution, inverse residual block, channel scaling and quantization to decompose the standard convolutional neural network into a lightweight model, as follows: Step 3.1.1, the machine learning module splits the standard convolution operation into depth-separable convolution and point-by-point convolution. Suppose the input feature map is ,in is the height, is the width, is the number of channels, Indicates size The set of real matrices of ; Depthwise separable convolution for each channel , the convolution kernel size is : ； in, represents the convolution operation, For the The convolution kernel of channels; For the The characteristics of each channel; For the The depth-wise separable convolution result of channels; Point-wise convolution uses The convolution kernel linearly combines the features of each channel: ； in, for The weight matrix of the convolution kernel; is the linear combination result of each channel feature; Step 3.1.2, the reverse residual block is divided into the following three steps: (1) Channel expansion: through Convolution expands the number of channels to : ； in is the expansion coefficient, used to increase the feature space; is the expanded feature map; Expand the matrix for the channel; Indicates size The set of real matrices of ; (2) Depth-wise separable convolution: for the expanded feature map conduct Depthwise separable convolution: ； in, is the depth convolution matrix; Indicates size The set of real matrices of ; for The depth-wise separable convolution result of ; (3) Channel contraction: through Convolution restores the number of channels to the number of output channels : ； in, Recover the matrix for the channel; Indicates size The set of real matrices of ; Step 3.1.3: Channel scaling by setting the scaling factor To reduce the number of channels per layer, set the number of channels to , then the number of channels after scaling for: ； Step 3.1.4: Quantization converts the model's weights and activations from floating point numbers to fixed point numbers, setting the original weight matrix to a floating point matrix , then the quantitative formula is: ； in, is the quantization scale factor, is the converted fixed-point matrix.

7. The fast path planning method for a lightweight UAV according to claim 5, characterized in that: The compressed neural network model is measured according to the UAV path planning business indicators described in step 3.2, as follows: Step 3.2.1: In terms of drone business optimization, smooth motion control is achieved and the speed of the drone is and angular velocity For smoothing, a first-order low-pass filter is used, and the formula is as follows: ； ； in, and is the filter coefficient, and its value range is ; , Speed , angular velocity No. Second filtering result; , Speed , angular velocity No. Second filtering result; , Speed , angular velocity Results without filtering; Step 3.2.

2. Loss function It includes two parts: mean square error MSE and binary cross entropy BCE: ； Among them, MSE is used to measure the error of the model when predicting continuous variables; BCE is used to deal with the loss of classification tasks. is the weight used to adjust the two parts of the loss.

8. The fast path planning method for a lightweight UAV according to claim 1, characterized in that: The strategy deployment module described in step 4 sends the path planning strategy generated by the machine learning module to the drone for path planning, as follows: Step 4.1, the strategy deployment module receives the path planning strategy generated by the machine learning module and parses it into low-level control instructions that the drone can recognize; Step 4.2, encoding the parsed control instructions into machine language or a set communication format; Step 4.3: The strategy deployment module sends instructions to the control module of the drone in real time according to the preset cycle or trigger condition. After sending, the strategy deployment module waits for the feedback signal from the drone to confirm whether the instruction is successfully executed. If it is executed, go to step 4.5, otherwise go to step 4.4; Step 4.4: If the drone shows in the feedback that it has failed to fully execute the command, the strategy deployment module adjusts and resends the correction command within the preset time; Step 4.5: If the UAV shows in the feedback that the command has been fully executed, the strategy deployment module continuously monitors the UAV's status information, including position, speed, and orientation, and returns the real-time environmental status to the machine learning module through a closed-loop feedback mechanism to update the model or re-plan the path.

9. The fast path planning method for a lightweight UAV according to claim 8, characterized in that: In step 4.2, the parsed control instructions are encoded into machine language or a specific communication format, including reduced instruction set, binary encoding compression, multi-instruction package transmission, and parameterized encoding design; The simplified instruction set is a simplified instruction set established for the commonly used control commands of drones; The binary code compression is to compress the instructions using binary code, mapping each control instruction to a unique binary code; The multi-command package transmission is to package multiple commands together and transmit them to the drone, reducing the frequency of command transmission one by one; The parameterized coding design adopts a parameterized coding mode to specify multiple pieces of information through one instruction.

10. A fast path planning device suitable for a lightweight unmanned aerial vehicle, characterized in that: The device is used to implement the fast path planning method applicable to lightweight unmanned aerial vehicles as described in any one of claims 1 to 9. The device includes a visual acquisition module, a machine learning module and a strategy deployment module. The functions of each module are as follows: The visual acquisition module collects and compresses the visual information images of the environment and transmits them to the machine learning module; The machine learning module first compresses the neural network model through deep separable convolution, inverse residual block, channel scaling and quantization processing, then measures the compressed neural network model according to the UAV path planning business indicators, and finally learns and tests the compressed neural network model to generate an optimized path planning strategy; The strategy deployment module sends the path planning strategy generated by the machine learning module to the drone for path planning.

Images