Obstacle recognition method of air purifier, electronic equipment and storage medium

Abstract

The invention relates to the technical field of air purifiers, and discloses an obstacle recognition method for an air purifier, electronic equipment and a storage medium, and the method comprises the following steps: S1, collecting distance information in an environment through a laser radar module, generating laser radar original data, capturing an environment image through an image collection module, generating image original data, and storing the image original data in a database; s2, preprocessing the original data of the laser radar, including filtering, noise reduction and coordinate transformation, and generating preprocessed laser radar data, S3, preprocessing the original data of the image; by constructing a multi-sensor collaborative fusion mechanism and dynamically optimizing a laser radar and visual data alignment strategy for different environments, the time-space consistency of point cloud data and image information is ensured, the problem of positioning deviation caused by sensor transmission delay can be eliminated, the accuracy of obstacle space coordinate mapping is ensured, and the accuracy of obstacle space coordinate mapping is improved. The navigation path planning error is further reduced; and the obstacle avoidance safety in a complex scene is improved.

Description

Technical Field

[0001] This invention relates to the field of air purifier technology, and in particular to an obstacle recognition method, electronic device and storage medium for an air purifier. Background Technology

[0002] Air purifiers, also known as air cleaners or air fresheners, are products that can adsorb, decompose, and transform various air pollutants to improve air cleanliness. They are primarily household and commercial air purifiers used to remove indoor air pollution. With industrial development, the amount of air pollutants is increasing, leading to a greater diversity of air purifier functions. Air purifiers use a high-voltage circuit to generate negative ions and static electricity, which, combined with a fan, circulates air, creating a filtration system. The fan inside the air purifier circulates indoor air, filtering out various pollutants and removing them, thus achieving the purpose of cleaning, dust removal, and purifying indoor air.

[0003] Currently, because air purifiers need to process LiDAR and visual data simultaneously in dynamic environments, the multi-sensor fusion system equipped with them cannot calibrate the timing misalignment between LiDAR scanning frames and camera acquisition frames in real time when performing real-time obstacle recognition. When there is a millisecond-level delay difference in sensor data transmission, it will cause spatial offset of the fused obstacle coordinates, resulting in increased obstacle avoidance path planning errors and making it impossible to guarantee the accuracy of obstacle positioning.

[0004] Therefore, an obstacle recognition method, electronic device, and storage medium for air purifiers are proposed to solve the above problems. Summary of the Invention

[0005] The main objective of this invention is to provide an obstacle recognition method, electronic device, and storage medium for an air purifier, in order to solve the problems mentioned in the background.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: an obstacle recognition method for an air purifier, comprising the following steps:

[0007] S1. Collect distance information in the environment through the lidar module to generate lidar raw data, and capture environmental images through the image acquisition module to generate image raw data;

[0008] S2. Preprocess the raw LiDAR data, including filtering, noise reduction and coordinate transformation, to generate preprocessed LiDAR data;

[0009] S3. Preprocess the original image data by sequentially performing histogram equalization, local brightness adaptive adjustment, and color correction to generate preprocessed image data.

[0010] S4. Input the preprocessed LiDAR data into the point cloud deep learning model to extract obstacle contour features, and simultaneously input the preprocessed image data into the convolutional neural network model to identify obstacle types.

[0011] S5. Integrate the obstacle contour features and obstacle type information to generate a comprehensive obstacle recognition result;

[0012] S6. Update the navigation map based on the comprehensive obstacle identification results, and adjust the operating path of the air purifier based on the dynamic path planning algorithm;

[0013] S7. Repeat steps S1 to S6 to achieve real-time obstacle avoidance.

[0014] Preferably, in step S2:

[0015] The filtering uses a Gaussian filtering algorithm, specifically including: setting the Gaussian kernel size to 3×3 pixels, the standard deviation σ=1.5, and performing a two-dimensional convolution operation on the raw LiDAR data to smooth the noise;

[0016] The noise reduction is achieved by statistical outlier removal, specifically by calculating the average distance between each point in the point cloud data and its 10 nearest neighbors. If the distance exceeds twice the standard deviation of the global average distance, it is identified as an outlier and deleted.

[0017] The coordinate transformation converts the lidar data from polar coordinates to Cartesian coordinates. The transformation formula is as follows:

[0018] ;

[0019] in Radial distance, representing the straight-line distance from the object to the sensor measured by the lidar. The scanning angle represents the emission angle of the lidar beam. and The coordinates are in the Cartesian coordinate system, representing the planar position of the object after transformation. The transformed point cloud data is used to construct a two-dimensional raster map with a resolution of 0.05m.

[0020] Preferably, in step S3, the local brightness adaptive adjustment includes the following sub-steps:

[0021] S31. Divide the image into 8×8 pixel local regions and calculate the average pixel brightness in each region. and standard deviation ;

[0022] S32, Based on preset target brightness parameters , Adjust the pixel values ​​using the following formula:

[0023] ;

[0024] in This represents the original brightness value of the pixel, indicating the grayscale intensity at coordinates (x, y), with a range of 0-255. This represents the average brightness of a local area, indicating the average brightness of all pixels within an 8×8 pixel region. The standard deviation of brightness in a local area represents the degree of dispersion of brightness values ​​within an 8×8 pixel region. The target average brightness is set at 128. The target brightness standard deviation is set to 52. The adjusted pixel brightness value;

[0025] S33. Perform color correction on the adjusted image: convert the RGB color space to HSV space, enhance the saturation in the S channel with an enhancement factor of 1.8, and finally output the preprocessed image data.

[0026] Preferably, in step S4:

[0027] The point cloud deep learning model adopts the multi-scale grouping structure of PointNet++, and its hierarchical feature extraction includes:

[0028] The first level samples 1024 key points with a search radius of 0.1m and MLP layer dimensions [64,128].

[0029] The second level of sampling includes 512 key points with a search radius of 0.3m and MLP layer dimensions of [128, 256].

[0030] The third level of sampling includes 256 key points with a search radius of 0.6m and an MLP layer dimension of [256, 512].

[0031] The convolutional neural network model adopts the ResNet-50 architecture. After removing the fully connected layers, the convolutional layers and residual blocks are retained. The last convolutional layer outputs a 512-dimensional feature vector.

[0032] The two models are executed synchronously through parallel computing on the GPU, with a point cloud data input frequency of 10Hz and an image data input frequency of 30Hz.

[0033] Preferably, in step S5, the feature fusion is achieved through channel splicing:

[0034] 512-dimensional point cloud feature vectors With 512-dimensional image feature vectors spliced ​​into 1024-dimensional fused features The fused data is input into a fully connected layer classifier, which consists of two 128-dimensional ReLU activation layers and a Softmax output layer, and outputs the probability distribution of obstacle categories.

[0035] Preferably, steps S4 and S5 are executed in a data processing unit that integrates an NVIDIA Jetson TX2 module, which includes a 256-core GPU and 8GB of LPDDR4 memory.

[0036] Accelerate model inference using the CUDA 10.2 parallel computing framework, compressing point cloud data processing latency to less than 25ms and image data processing latency to less than 35ms;

[0037] Data transmission uses a PCIe 3.0×4 bus with a bandwidth of 32Gbps.

[0038] Preferably, in step S6, the dynamic path planning algorithm employs an improved A* algorithm:

[0039] The basic cost function is:

[0040] ;

[0041] in The actual cost from the starting point to node n. For heuristic cost estimation;

[0042] When identifying dynamic obstacles, a risk factor is added to the cost function:

[0043]

[0044] in Let be the original cost function value of node n, representing the cumulative cost of node n in path planning. Risk coefficient, default value 5.0, is a dimensionless parameter that adjusts the weight of the impact of obstacles. Let n be the minimum distance between node n and the obstacle, representing the safety margin. The adjusted cost function value is used to optimize the path by considering risk factors.

[0045] The path replanning is triggered when an obstacle intrudes into the safe radius of the purifier's trajectory. The safe radius is set to 1.2 times the diameter of the purifier.

[0046] Preferably, the dynamic obstacle recognition logic includes:

[0047] I. Track the obstacle position on three consecutive frames and calculate the centroid displacement:

[0048] ;

[0049] in This represents the planar distance that a dynamic obstacle travels between two adjacent image frames. The horizontal pixel coordinates of the centroid of the dynamic obstacle in the current frame image. Here are the horizontal pixel coordinates of the centroid of the dynamic obstacle in the previous frame image. Here are the vertical pixel coordinates of the centroid of the dynamic obstacle in the current frame image. t represents the vertical pixel coordinates of the centroid of the dynamic obstacle in the previous frame image, and t represents the timestamp of image acquisition.

[0050] II. When If m and last for 2 cycles, it is determined to be a dynamic obstacle;

[0051] III. The classification of dynamic obstacles includes: pet cats: movement speed 0.5-2m / s, electric toy cars: speed 0.3-1m / s, and walking children: speed 0.4-1.5m / s. The classification confidence threshold is set at 85%.

[0052] Preferably, the electronic device comprises:

[0053] LiDAR module: Employs a 905nm wavelength ToF sensor, with a 360° scanning angle, 0.5° angular resolution, and a detection range of 0.1-10m;

[0054] Image acquisition module: Equipped with a Sony IMX477 camera, 1920×1080 resolution, 60fps frame rate, and built-in HDR mode;

[0055] Data processing unit: integrates an NVIDIA Jetson TX2 module, a built-in Pascal architecture GPU, and is equipped with 32GB eMMC memory to execute an obstacle recognition method for an air purifier;

[0056] Navigation control module: Based on STM32F746 microcontroller, it receives obstacle coordinates in real time and generates PWM motor drive signals;

[0057] High-speed data bus: Adopting a two-layer communication architecture, the LiDAR is transmitted via UART@115200bps, image data is transmitted via MIPI CSI-2 channel, and control commands are transmitted via CAN bus, with a system response delay of less than 100ms.

[0058] Preferably, the storage medium uses a UFS 3.1 standard flash memory chip and stores a computer program. When the program is executed by a processor, it implements an obstacle recognition method for an air purifier, including:

[0059] I. LiDAR data acquisition and preprocessing procedures;

[0060] II. Image Adaptive Enhancement and Feature Extraction Program;

[0061] III. Point cloud and image dual-stream fusion classification program;

[0062] IV. Dynamic path planning decision-making process;

[0063] V. Equipment control command generation program;

[0064] All program code runs using a circular buffer mechanism, with memory usage less than 1.5GB.

[0065] The present invention has the following beneficial effects:

[0066] 1. In this invention, when performing obstacle recognition for air purifiers, a multi-sensor collaborative fusion mechanism is constructed, and a dynamic optimization strategy for aligning LiDAR and visual data for different environments is adopted to ensure the temporal and spatial consistency of point cloud data and image information. This eliminates the positioning deviation problem caused by sensor transmission delay, ensures the accuracy of obstacle spatial coordinate mapping, further reduces navigation path planning errors, and improves obstacle avoidance safety in complex scenarios.

[0067] 2. In this invention, when performing dynamic obstacle recognition, a risk factor is added to the cost function by improving the A* algorithm, the minimum distance to the obstacle is calculated in real time and the path weight is dynamically adjusted. At the same time, a safety radius triggering replanning mechanism is set, so that the system can actively avoid the risk of dynamic obstacle intrusion, ensure the safety margin of the air purifier's operating path in the scenario of moving obstacles, and improve the reliability of real-time obstacle avoidance decision-making.

[0068] 3. In the data processing architecture design of this invention, the parallel acceleration of point cloud and image dual-model inference is achieved by GPU, and the layered bus architecture is combined to realize millisecond-level data transmission. At the same time, a circular buffer mechanism is adopted to control memory usage, so that the system can still maintain millisecond-level full-process response latency under the condition of integrating deep learning models, ensuring the efficient execution and real-time performance of complex algorithms in embedded devices. Attached Figure Description

[0069] Figure 1 This is a flowchart of an obstacle recognition method for an air purifier according to the present invention. Detailed Implementation

[0070] 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.

[0071] For specific implementation examples, please refer to: Figure 1 An obstacle recognition method for an air purifier includes the following steps:

[0072] S1. Collect distance information in the environment through the lidar module to generate lidar raw data, and capture environmental images through the image acquisition module to generate image raw data;

[0073] S2. Preprocess the raw lidar data, including filtering, noise reduction and coordinate transformation, to generate preprocessed lidar data;

[0074] S3. Preprocess the original image data by sequentially performing histogram equalization, local brightness adaptive adjustment and color correction to generate preprocessed image data.

[0075] S4. Input the preprocessed LiDAR data into the point cloud deep learning model to extract obstacle contour features, and simultaneously input the preprocessed image data into the convolutional neural network model to identify obstacle types.

[0076] S5. Integrate obstacle contour features and obstacle type information to generate a comprehensive obstacle recognition result;

[0077] S6. Update the navigation map based on the comprehensive obstacle recognition results, and adjust the air purifier's operating path based on the dynamic path planning algorithm;

[0078] S7. Repeat steps S1 to S6 to achieve real-time obstacle avoidance.

[0079] In step S2:

[0080] The filtering uses a Gaussian filtering algorithm, which specifically includes: setting the Gaussian kernel size to 3×3 pixels and the standard deviation σ=1.5, and performing a two-dimensional convolution operation on the raw LiDAR data to smooth the noise;

[0081] Noise reduction is achieved by statistical outlier removal, specifically by calculating the number of outliers for each point in the point cloud data. its nearest neighbor Average distance of points ,when Then it is identified as an outlier and deleted, where:

[0082] ;

[0083] in Let the three-dimensional coordinates of the point cloud to be detected be... The three-dimensional coordinates of the nearest neighbor point. This is the number of nearest neighbors; the default value is 10. The average distance of the global point cloud. The standard deviation of the global point cloud distance. This is the outlier determination coefficient, with a default value of 2.0.

[0084] Coordinate transformation converts LiDAR data from polar coordinates to Cartesian coordinates. The transformation formula is:

[0085] ;

[0086] in Radial distance, representing the straight-line distance from the object to the sensor measured by the lidar. The scanning angle represents the emission angle of the lidar beam. and The coordinates are in the Cartesian coordinate system, representing the planar position of the object after transformation. The transformed point cloud data is used to construct a two-dimensional raster map with a resolution of 0.05m.

[0087] In step S3, the local brightness adaptive adjustment includes the following sub-steps:

[0088] S31. Divide the image into 8×8 pixel local regions and calculate the average pixel brightness in each region. and standard deviation ;

[0089] S32, Based on preset target brightness parameters , Adjust the pixel values ​​using the following formula:

[0090] ;

[0091] in This represents the original brightness value of the pixel, indicating the grayscale intensity at coordinates (x, y), with a range of 0-255. This represents the average brightness of a local area, indicating the average brightness of all pixels within an 8×8 pixel region. The standard deviation of brightness in a local area represents the degree of dispersion of brightness values ​​within an 8×8 pixel region. The target average brightness is set at 128. The target brightness standard deviation is set to 52. The adjusted pixel brightness value;

[0092] S33. Perform color correction on the adjusted image: convert the RGB color space to HSV space, enhance the saturation in the S channel with an enhancement factor of 1.8, and finally output the preprocessed image data.

[0093] In step S4:

[0094] The point cloud deep learning model adopts the multi-scale grouping structure of PointNet++, and its hierarchical feature extraction includes:

[0095] The first level samples 1024 key points with a search radius of 0.1m and MLP layer dimensions [64,128].

[0096] The second level of sampling includes 512 key points with a search radius of 0.3m and MLP layer dimensions of [128, 256].

[0097] The third level of sampling includes 256 key points with a search radius of 0.6m and an MLP layer dimension of [256, 512].

[0098] The convolutional neural network model uses the ResNet-50 architecture, which removes the fully connected layers but retains the convolutional layers and residual blocks. The last convolutional layer outputs a 512-dimensional feature vector.

[0099] The two models are executed synchronously through parallel computing on the GPU, with a point cloud data input frequency of 10Hz and an image data input frequency of 30Hz.

[0100] In step S5, feature fusion is achieved through channel splicing:

[0101] 512-dimensional point cloud feature vectors With 512-dimensional image feature vectors spliced ​​into 1024-dimensional fused features The fused data is then input into a fully connected layer classifier, which consists of two 128-dimensional ReLU activation layers and a Softmax output layer. The output layer shows the probability distribution of obstacle categories, specifically calculated using the Softmax function.

[0102] ;

[0103] in This represents the total number of obstacle categories, including furniture, pets, and people. , For category The corresponding weight vector and bias term, It is a 1024-dimensional fused feature vector. For input features to belong to categories The probability of;

[0104] Steps S4 and S5 are executed in the data processing unit, which integrates an NVIDIA Jetson TX2 module containing a 256-core GPU and 8GB of LPDDR4 memory.

[0105] Accelerate model inference using the CUDA 10.2 parallel computing framework, compressing point cloud data processing latency to less than 25ms and image data processing latency to less than 35ms;

[0106] Data transmission uses a PCIe 3.0×4 bus with a bandwidth of 32Gbps.

[0107] In step S6, the dynamic path planning algorithm uses an improved A* algorithm:

[0108] The basic cost function is:

[0109] ;

[0110] in The actual cost from the starting point to node n. For heuristic cost estimation;

[0111] When identifying dynamic obstacles, a risk factor is added to the cost function:

[0112] ;

[0113] in Let be the original cost function value of node n, representing the cumulative cost of node n in path planning. Risk coefficient, default value 5.0, is a dimensionless parameter that adjusts the weight of the impact of obstacles. Let n be the minimum distance between node n and the obstacle, representing the safety margin. The adjusted cost function value is used to optimize the path by considering risk factors.

[0114] The path replanning is triggered when an obstacle intrudes into the safe radius of the purifier's trajectory. The safe radius is set to 1.2 times the diameter of the purifier.

[0115] The dynamic obstacle recognition logic includes:

[0116] I. Track the obstacle position on three consecutive frames and calculate the centroid displacement. :

[0117] ;

[0118] in This represents the planar distance that a dynamic obstacle travels between two adjacent image frames. The horizontal pixel coordinates of the centroid of the dynamic obstacle in the current frame image. Here are the horizontal pixel coordinates of the centroid of the dynamic obstacle in the previous frame image. Here are the vertical pixel coordinates of the centroid of the dynamic obstacle in the current frame image. t represents the vertical pixel coordinates of the centroid of the dynamic obstacle in the previous frame image, and t represents the timestamp of image acquisition.

[0119] II. Derivation of real-time speed:

[0120] in ;

[0121] in Instantaneous velocity This is the frame interval, defaulting to 33ms;

[0122] III. When If m and last for 2 cycles, it is determined to be a dynamic obstacle;

[0123] IV. The classification of dynamic obstacles includes: pet cats: movement speed 0.5-2m / s, electric toy cars: speed 0.3-1m / s, and walking children: speed 0.4-1.5m / s. The classification confidence threshold is set at 85%.

[0124] The electronic device includes:

[0125] LiDAR module: Employs a 905nm wavelength ToF sensor, with a 360° scanning angle and 0.5° angular resolution, operating under ambient light intensity. Under interference, the signal-to-noise ratio is maintained through adaptive transmit power adjustment:

[0126] ;

[0127] Where SNR is the signal-to-noise ratio of the lidar, with a threshold greater than 8.0. The laser emission power is dynamically adjusted to 2mW. The target reflectivity, the default value is 0.3. This is the atmospheric attenuation coefficient, with a default value of 0.02. , For the target distance, For ambient light intensity, This refers to the photosensitive area of ​​the detector, which is 1.2m² by default. , The circuit noise floor is set to 0.05μW by default.

[0128] when When it is less than 0.8, the control chip presses... Increase power;

[0129] Image acquisition module: Equipped with a Sony IMX477 camera, 1920×1080 resolution, 60fps frame rate, and built-in HDR mode;

[0130] Data processing unit: integrates an NVIDIA Jetson TX2 module, a built-in Pascal architecture GPU, and is equipped with 32GB eMMC memory to execute an obstacle recognition method for an air purifier;

[0131] Navigation control module: Based on STM32F746 microcontroller, it receives obstacle coordinates in real time and generates PWM motor drive signals;

[0132] High-speed data bus: Adopting a two-layer communication architecture, the LiDAR is transmitted via UART@115200bps, image data is transmitted via MIPI CSI-2 channel, and control commands are transmitted via CAN bus, with a system response delay of less than 100ms.

[0133] The storage medium uses UFS 3.1 standard flash memory chips and stores a computer program. When the program is executed by the processor, it implements an obstacle recognition method for an air purifier, including:

[0134] I. LiDAR data acquisition and preprocessing procedures;

[0135] II. Image Adaptive Enhancement and Feature Extraction Program;

[0136] III. Point cloud and image dual-stream fusion classification program;

[0137] IV. Dynamic path planning decision-making process;

[0138] V. Equipment control command generation program;

[0139] All program code runs using a circular buffer mechanism, with memory usage less than 1.5GB, satisfying the following:

[0140] ;

[0141] in This represents the actual memory usage. This is the size of the circular buffer; the default is 1.2GB. This is the compression factor, with a default value of 0.4. This represents the system's basic memory overhead, which is 0.3GB by default.

[0142] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An obstacle recognition method for an air purifier, characterized in that, Includes the following steps: S1. Collect distance information in the environment through the lidar module to generate lidar raw data, and capture environmental images through the image acquisition module to generate image raw data; S2. Preprocess the raw LiDAR data, including filtering, noise reduction and coordinate transformation, to generate preprocessed LiDAR data; S3. Preprocess the original image data by sequentially performing histogram equalization, local brightness adaptive adjustment, and color correction to generate preprocessed image data. S4. Input the preprocessed LiDAR data into the point cloud deep learning model to extract obstacle contour features, and simultaneously input the preprocessed image data into the convolutional neural network model to identify obstacle types. S5. Integrate the obstacle contour features and obstacle type information to generate a comprehensive obstacle recognition result; S6. Update the navigation map based on the comprehensive obstacle identification results, and adjust the operating path of the air purifier based on the dynamic path planning algorithm; S7. Repeat steps S1 to S6 to achieve real-time obstacle avoidance.

2. The obstacle recognition method for an air purifier according to claim 1, characterized in that, In step S2: The filtering uses a Gaussian filtering algorithm, specifically including: setting the Gaussian kernel size to 3×3 pixels, the standard deviation σ=1.5, and performing a two-dimensional convolution operation on the raw LiDAR data to smooth the noise; The noise reduction is achieved by statistical outlier removal, specifically: calculating the average distance between each point in the point cloud data and its 10 nearest neighbors, and determining an outlier as a point when the distance exceeds twice the standard deviation of the global average distance; The coordinate transformation converts the lidar data from polar coordinates to Cartesian coordinates. The transformation formula is as follows: ； in Radial distance, representing the straight-line distance from the object to the sensor measured by the lidar. The scanning angle represents the emission angle of the lidar beam. and The coordinates are in the Cartesian coordinate system, representing the planar position of the object after transformation. The transformed point cloud data is used to construct a two-dimensional raster map with a resolution of 0.05m.

3. The obstacle recognition method for an air purifier according to claim 1, characterized in that, In step S3, the local brightness adaptive adjustment includes the following sub-steps: S31. Divide the image into 8×8 pixel local regions and calculate the average pixel brightness in each region. and standard deviation ; S32, Based on preset target brightness parameters , Adjust the pixel values ​​using the following formula: ； in This represents the original brightness value of the pixel, indicating the grayscale intensity at coordinates (x, y), with a range of 0-255. This represents the average brightness of a local area, indicating the average brightness of all pixels within an 8×8 pixel region. The standard deviation of brightness in a local area represents the degree of dispersion of brightness values ​​within an 8×8 pixel region. The target average brightness is set at 128. The target brightness standard deviation is set to 52. The adjusted pixel brightness value; S33. Perform color correction on the adjusted image: convert the RGB color space to HSV space, enhance the saturation in the S channel with an enhancement factor of 1.8, and finally output the preprocessed image data.

4. The obstacle recognition method for an air purifier according to claim 1, characterized in that, In step S4: The point cloud deep learning model adopts the multi-scale grouping structure of PointNet++, and its hierarchical feature extraction includes: The first level samples 1024 key points with a search radius of 0.1m and MLP layer dimensions [64,128]. The second level of sampling includes 512 key points with a search radius of 0.3m and MLP layer dimensions of [128, 256]. The third level of sampling includes 256 key points with a search radius of 0.6m and an MLP layer dimension of [256, 512]. The convolutional neural network model adopts the ResNet-50 architecture. After removing the fully connected layers, the convolutional layers and residual blocks are retained. The last convolutional layer outputs a 512-dimensional feature vector. The two models are executed synchronously through parallel computing on the GPU, with a point cloud data input frequency of 10Hz and an image data input frequency of 30Hz.

5. The obstacle recognition method for an air purifier according to claim 1, characterized in that, In step S5, the feature fusion is achieved through channel splicing: 512-dimensional point cloud feature vectors With 512-dimensional image feature vectors spliced ​​into 1024-dimensional fused features The fused data is input into a fully connected layer classifier, which consists of two 128-dimensional ReLU activation layers and a Softmax output layer, and outputs the probability distribution of obstacle categories.

6. The obstacle recognition method for an air purifier according to claim 1, characterized in that: Steps S4 and S5 are executed in the data processing unit, which integrates an NVIDIA Jetson TX2 module containing a 256-core GPU and 8GB of LPDDR4 memory. Accelerate model inference using the CUDA 10.2 parallel computing framework, compressing point cloud data processing latency to less than 25ms and image data processing latency to less than 35ms; Data transmission uses a PCIe 3.0×4 bus with a bandwidth of 32Gbps.

7. The obstacle recognition method for an air purifier according to claim 1, characterized in that, In step S6, the dynamic path planning algorithm employs an improved A* algorithm: The basic cost function is: ； in The actual cost from the starting point to node n. For heuristic cost estimation; When identifying dynamic obstacles, a risk factor is added to the cost function: ； in Let be the original cost function value of node n, representing the cumulative cost of node n in path planning. Risk coefficient, default value 5.0, is a dimensionless parameter that adjusts the weight of the impact of obstacles. Let n be the minimum distance between node n and the obstacle, representing the safety margin. The adjusted cost function value is used to optimize the path by considering risk factors. The path replanning is triggered when an obstacle intrudes into the safe radius of the purifier's trajectory. The safe radius is set to 1.2 times the diameter of the purifier.

8. The obstacle recognition method for an air purifier according to claim 7, characterized in that, The dynamic obstacle recognition logic includes: I. Track the obstacle position on three consecutive frames and calculate the centroid displacement: ； in This represents the planar distance that a dynamic obstacle travels between two adjacent image frames. The horizontal pixel coordinates of the centroid of the dynamic obstacle in the current frame image. Here are the horizontal pixel coordinates of the centroid of the dynamic obstacle in the previous frame image. Here are the vertical pixel coordinates of the centroid of the dynamic obstacle in the current frame image. t represents the vertical pixel coordinates of the centroid of the dynamic obstacle in the previous frame image, and t represents the timestamp of image acquisition. II. When If m and last for 2 cycles, it is determined to be a dynamic obstacle; III. The classification of dynamic obstacles includes: pet cats: movement speed 0.5-2m / s, electric toy cars: speed 0.3-1m / s, and walking children: speed 0.4-1.5m / s. The classification confidence threshold is set at 85%.

9. An electronic device for an air purifier, characterized in that, The electronic device includes: LiDAR module: Employs a 905nm wavelength ToF sensor, with a 360° scanning angle, 0.5° angular resolution, and a detection range of 0.1-10m; Image acquisition module: Equipped with a Sony IMX477 camera, 1920×1080 resolution, 60fps frame rate, and built-in HDR mode; Data processing unit: integrates an NVIDIA Jetson TX2 module, a built-in Pascal architecture GPU, and is equipped with 32GB eMMC memory, for executing the obstacle recognition method of any one of the air purifiers in claims 1-8; Navigation control module: Based on STM32F746 microcontroller, it receives obstacle coordinates in real time and generates PWM motor drive signals; High-speed data bus: Adopting a two-layer communication architecture, the LiDAR is transmitted via UART@115200bps, image data is transmitted via MIPI CSI-2 channel, and control commands are transmitted via CAN bus, with a system response delay of less than 100ms.

10. A storage medium for an air purifier, characterized in that, The storage medium uses a UFS 3.1 standard flash memory chip and stores a computer program. When the program is executed by the processor, it implements the obstacle recognition method for any one of claims 1-8 for an air purifier, comprising: I. LiDAR data acquisition and preprocessing procedures; II. Image Adaptive Enhancement and Feature Extraction Program; III. Point cloud and image dual-stream fusion classification program; IV. Dynamic path planning decision-making process; V. Equipment control command generation program; All program code runs using a circular buffer mechanism, with memory usage less than 1.5GB.

Images