CN113989718A - Human body target detection method facing radar signal heat map - Google Patents
Human body target detection method facing radar signal heat map Download PDFInfo
- Publication number
- CN113989718A CN113989718A CN202111273690.4A CN202111273690A CN113989718A CN 113989718 A CN113989718 A CN 113989718A CN 202111273690 A CN202111273690 A CN 202111273690A CN 113989718 A CN113989718 A CN 113989718A
- Authority
- CN
- China
- Prior art keywords
- data
- network
- heat map
- radar signal
- matrix
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The human body target detection method is based on a residual error network ResNet and adopts a radar signal heat map as network input. Compared with video images, the heat map can display target information from horizontal and vertical bidirectional visual angles, effectively removes redundant backgrounds and is more visual. In order to simplify the training process, the method preprocesses the heat map input, integrates network input parameters and recombines heat map image data, and is also beneficial to improving the training efficiency and reducing the training time. In a training model, the method extracts features output by a residual error network, and fuses horizontal and vertical heat map features by means of feature matrix reconstruction operation at the end of the network so as to further improve the feature extraction effect.
Description
Technical Field
The invention belongs to the field of target image identification, and particularly relates to a human body target detection method facing a radar signal heat map.
Background
The radar signal heat map, that is, the radio frequency heat map obtained by the radio frequency transceiver such as radar, is adopted by various detection systems due to its high distance resolution and low obtaining difficulty, and is widely used in the fields of military affairs, traffic, geological engineering, etc. Extracting and identifying various targets, especially human targets, from radar signal heat maps is one of the basic requirements and important tasks of various detection systems.
Human target detection has been a research hotspot in the field of computer vision. In early studies, the target detection method usually needs to extract image features, and a dpm (deformable Parts model) model is used to predict a target frame (Bounding Box) with a higher score in a sliding window manner. This approach is time consuming and not very accurate. And then, the researchers introduce deep learning into the field of target detection, so that the performance of target detection is greatly improved. Such methods can be divided into two categories. One is to divide the object identification and object positioning into Two steps to be respectively completed, namely Two-stage, such as R-CNN, Fast R-CNN, FasterR-CNN, etc., which has low identification error rate but slow speed, thus being not suitable for real-time detection scenes. The other is One-stage, which abandons the region proposal method (RegionProposal) and directly generates the target coordinate values and classification probabilities, typical methods include Yolo, SSD, Yolo v2, etc. The identification speed is improved, certain defects still exist in the aspects of multi-scale prediction, basic classification networks and the like, and the overall accuracy under a complex detection scene is not ideal enough. These studies are mainly directed to conventional video detection, and cannot be directly applied to human target detection in radar signal heat maps. In recent years, colleges and universities typified by the Massachusetts Institute of Technology (MIT) have actively pursued this field. In 2017, Adib et al developed a radio frequency signal based RF-Capture system that could Capture human objects behind a wall through a wall. In 2018, Zhao and the like develop an RF-Pose system for human body target detection and key point positioning work by utilizing WiFi wireless signals on the basis of RF-Capture. These pioneering studies have made good attempts to detect human targets in radar signal heatmaps, but there are many limitations, such as RF-Capture is limited to a single action performed by a single person (i.e. walking towards the device), and RF-pos is only applicable to a few people's scenes. In general, such methods have not been able to achieve high accuracy in complex environments.
Disclosure of Invention
Aiming at the problems, the invention provides a human body target detection method facing a radar signal heat map, which simultaneously uses horizontal and vertical heat map input and introduces a residual error network ResNet (ResidualNet) to improve the identification performance. The method is based on a residual error network ResNet and adopts a radar signal heat map as network input. Compared with video images, the heat map can display target information from horizontal and vertical bidirectional visual angles, effectively removes redundant backgrounds and is more visual. In order to simplify the training process, the method preprocesses the heat map input, integrates network input parameters and recombines heat map image data, and is also beneficial to improving the training efficiency and reducing the training time. In a training model, the method extracts features output by a residual error network, and fuses horizontal and vertical heat map features by means of feature matrix reconstruction operation at the end of the network so as to further improve the feature extraction effect.
A human body target detection method facing a radar signal heat map comprises the following steps:
step 1: preprocessing data;the data comprises radio frequency data and video data, the radio frequency data is radio frequency signals sent by a multi-antenna array, the signals contact a target in a target area and are reflected back to an original point, vertical and horizontal bidirectional target space information is obtained according to the reflected signals, signal heat map original data is constructed, and then noise reduction, cutting and marking are carried out; the original data is a numerical matrix generated after each frame of radar signal is directly acquired, wherein the 2 nd, 4 th, 6 th and 8 th dimensions represent horizontal direction data, the 1 st, 3 th, 5 th and 7 th dimensions represent vertical direction data, the horizontal direction data, the 3 th, the 5 th and the 7 th dimensions represent vertical direction data, the vertical direction data are respectively taken out and then processed by image Fourier transform to obtain a heat map matrix, and a horizontal heat map data set and a vertical heat map data set are formed; the video data is obtained by synchronously recording a target scene by using a depth camera, and performing target identification storage and human body diagram extraction on a recorded video image by using a Faster R-CNN network to construct a label set for subsequent network training;1
step 2: constructing a feature extraction network based on ResNet-34; according to a residual block structure in the ResNet network, constructing a feature extraction network: taking a network structure of ResNet-34 as a main network, selecting a Basic Block type as a residual Block, setting the number set of the residual blocks as [3,4,6,3], wherein the numerical value represents the number of the residual blocks contained in the Layer of each residual network; the network is used for extracting a characteristic matrix for subsequent network processing;
and step 3: reconstructing a characteristic matrix; establishing a subsequent network architecture after the characteristic extraction network, and reconstructing the characteristic matrix through matrix flattening, deformation operation and a full connection layer to perfect the characteristic matrix;
and 4, step 4: fusing horizontal and vertical heat map features; performing fusion operation on the feature matrix obtained in the step; because the heat map is divided into a horizontal data block and a vertical data block, a block diagram data matrix is obtained after the features of the two heat maps are fused; an additional network architecture which is constructed by the step 3 and the step 4 is called a post-network and is used for processing the output of the feature extraction network;
and 5: constructing a target detection frame; and performing image segmentation operation on the original image by using a block diagram data matrix output by a post-network so as to realize the visual effect of human body target detection.
Further, the method comprises the steps of: in the step 1, label data are generated by video data through a Fast R-CNN network and are subjected to batch processing operation, so that the target detection regression frame coordinates of each original image are extracted, and finally, the regression frame coordinate data are converted into a Tensor data type through a DataLoader method of a Pyroch, so that a label set is obtained.
Further, in step 1, white noise removal is required to be performed on the original data of the signal heat map to complete noise reduction processing, the white noise is generally a numerical value with data close to 0 or lower than the surrounding, and the heat map data is converted into a numerical matrix after the operations are completed; in order to complete the cutting and marking work of the original data, cutting and compressing the numerical matrix generated by each data frame to make the numerical matrix suitable for the network input of the step 2; and the matrix representing the horizontal data is marked and distinguished from the matrix representing the vertical data, respectively. And finally, dividing the processed data into a training set and a test set according to the ratio of 8: 2.
Further, in step 1, for the generated tag set, training set and test set, data encapsulation is performed by using a Pytorch machine learning library, and finally, an encapsulated DataLoader is generated to be loaded on a network.
Further, in step 2, the feature extraction takes four layers as a main framework and is assisted by an upper pooling Layer and a lower pooling Layer; for network input, firstly, a first convolutional Layer and a pooling Layer are used for scaling a feature map, and then the feature map enters a main unit formed by four layers, each Layer consists of a plurality of residual blocks, and every two convolutional layers in the main unit form a residual block structure in a mapping connection mode; for the network output, processing is performed through a global average pooling layer.
Further, in step 2, the type of residual Block used in the present network is Basic Block, and each residual Block is composed of two 3 × 3 convolutional layers.
Further, in step 3, for the horizontal heat map and the vertical heat map, two feature matrices Out1 and Out2 may be obtained through step 2, respectively, and the feature matrices are reconstructed; firstly, performing Flatten operation on multidimensional data, and then performing characteristic processing on a full connection layer FC of 5x4 to achieve matrix optimization; wherein, the parameter of Flatten needs to be changed to adapt to different lengths and widths of network output; after the full connection layer FC, converting the characteristic matrixes of the horizontal heat map and the vertical heat map into final network output; and adopting Reshape operation after the full connection layer for the output display of the test set.
Further, in step 4, the fusion operation includes two different fusion types, namely Concat and Add; add fusion operations are used to accomplish horizontal, vertical data fusion.
Further, in step 5, the visual effect is output according to each frame of image; the background of the image is a two-dimensional plane of an original radar signal range, and then the position of the background of the image is marked according to coordinate set data of each regression frame, so that the effect display of human body target block diagram identification is realized.
Further, in step 5, the two-dimensional plane of the original radar signal range is defined as a null plane in which depth information is cut out from the arc-shaped area scanned by the radar signal.
The invention has the beneficial effects that:
(1) and constructing a feature extraction network taking the residual error network as a main framework. The method is beneficial to solving the degradation problem of the deep network and improving the characteristic extraction effect. Meanwhile, the combination of the deep neural network and the residual block is equivalent to the simplified equivalent mapping, additional parameters are not generated, the computational complexity is not increased, and the end-to-end back propagation training can still be realized.
(2) And adding a post structure of the feature extraction network. The post structure is the post network mentioned above, on one hand, the optimization and the perfection of the characteristic matrix are facilitated, and the accuracy rate of the final block diagram identification is also improved; on the other hand, the reconstruction of the matrix enables the fusion of the horizontal and vertical heat map characteristics to be more convenient and faster, and the network training time is shortened to a certain extent.
Drawings
Fig. 1 is a flow chart of data preprocessing according to an embodiment of the present invention.
Fig. 2 is a schematic diagram of a feature extraction network model architecture according to an embodiment of the present invention.
Fig. 3 is a schematic diagram of a post-network model architecture according to an embodiment of the present invention.
Detailed Description
The technical scheme of the invention is further explained in detail by combining the drawings in the specification.
The invention provides a target detection method facing a radar signal heat map. The method is based on a residual error network ResNet and adopts a radar signal heat map as network input. Compared with video images, the heat map can display target information from horizontal and vertical bidirectional visual angles, effectively removes redundant backgrounds and is more visual. In order to simplify the training process, the method preprocesses the heat map input, integrates network input parameters and recombines heat map image data, and is also beneficial to improving the training efficiency and reducing the training time. In a training model, the method extracts features output by a residual error network, and fuses horizontal and vertical heat map features by means of feature matrix reconstruction operation at the end of the network so as to further improve the feature extraction effect.
The target detection method facing the radar signal heat map specifically comprises the following steps:
step 1 (data preprocessing): the raw data in the present invention includes two types. One type is radio frequency data, which is radio frequency signals emitted by a multi-antenna array, the signals contact a target in a target area and are reflected back to an original point, and vertical and horizontal bidirectional target space information is obtained according to the reflected signals, so that a Fast Fourier Transform (FFT) signal heat map is constructed. The other type is video data, which is a video obtained by synchronously recording a target scene by using a depth camera. For these two types of data, the present invention will perform the preprocessing shown in fig. 1, and the specific steps are as follows.
(1) As shown in the left branch of fig. 1, label data is generated from the original video data through Fast R-CNN network and batch processing operation is performed, so as to extract the coordinates of the target detection regression frame of each original image, and finally generate a label set.
(2) As shown in the right branch of fig. 1, the original heat map is subjected to noise reduction processing, the heat map data is converted into a numerical matrix, a data set suitable for network input is generated through matrix adjustment operation and channel conversion, and finally the data set is segmented into a training set and a test set.
(3) And for the label set, the training set and the test set generated in the steps, carrying out data encapsulation by using a Pythrch machine learning library, and finally generating an encapsulated DataLoader to be loaded in a network.
Step 2 (feature extraction network based on ResNet-34): as shown in fig. 2, the overall network architecture is optimized according to the infrastructure of the ResNet 34. The network takes four layers as a main framework and is assisted by an upper pooling Layer and a lower pooling Layer. For network input, scaling of the profile is first performed using the first convolutional Layer (7x7) and the pooling Layer, and then into the master unit consisting of four layers. In fig. 2, four layers are distinguished according to different gray levels and color depths, and convolution layers with the same color depth belong to the same Layer. Each Layer consists of several residual blocks. The type of the residual Block used by the network is Basic Block, each residual Block is composed of two convolution layers of 3x3, in fig. 2, the two convolution layers connected by the dotted line form a residual Block structure, each dotted line in the diagram represents the mapping connection of each residual Block, and the mapping mode can reversely propagate the gradient to a shallower network layer, thereby being beneficial to constructing a deeper network architecture. The step size of the first residual block of the second, third and fourth layers is 2, and the other step sizes are 1, which are used for training in order to narrow down features in the deep network. After the network architecture of four layers, the feature matrix is reduced to 32 times of the original image matrix, thereby completing the feature extraction operation. Finally, the whole network is regularized in structure by global average pooling Layer (GlobalAverage Pool Layer) processing to prevent the overfitting problem. Here, the network architecture of the original ResNet34 is terminated by a full connection layer FC, the present invention eliminates the original full connection layer FC, and then connects a post network PostNetwork respectively for the network inputs of the horizontal and vertical heat maps to perform the feature map optimization processing, as described in step 3 and step 4.
Step 3 (reconstructing feature matrix): for the horizontal heat map and the vertical heat map, two feature matrices Out1, Out2 can be obtained by step 2, respectively, and they will be subjected to the reconstruction process shown in fig. 3. The matrix optimization is achieved by first performing a Flatten operation on the multidimensional data and then performing feature processing on the full link layer FC of 5x 4. Wherein, the parameter of Flatten needs to be changed to adapt to different length and width of the network output. After the full connectivity layer FC, the feature matrices of the horizontal and vertical heatmaps have been converted to the final network output. Since the maximum number of detection targets for target detection is set to 5, and each block coordinate set is composed of two coordinates (i.e., 4 numerical values), the parameter of the full link layer is 5 × 4. Finally, in order to keep the data formats of the network output and the tag set consistent, a Reshape operation is adopted for the test set output display.
Step 4 (fusion horizontal, vertical heat map features): as shown in fig. 3, after the horizontal heat map and the vertical heat map data are processed in step 3, a feature fusion operation is performed to obtain a final network output. The fusion operation includes two different fusion types, i.e., Concat and Add. The invention adopts Add fusion operation to complete horizontal and vertical data fusion, namely, each element of each dimension of the horizontal and vertical heat maps is added without changing the size of the element; in contrast, the Concat operation merges some dimension of the two sets of data to change the output size. Therefore, the PostNetwork Postnetwork formed by the steps 3 and 4 is constructed, and the feature extraction network is combined with the network architecture of the PostNetwork to form the whole network system.
Step 5 (constructing a target detection box): for the network output of step 4, the feature matrix has been converted into regression frame coordinate data for human detection due to the existence of the full connection layer FC in step 3. Through visual display, each image of the test set can be directly subjected to image processing so as to display all human body regression block diagrams in each image and mark the human body regression block diagrams in the images. In practical application, the output of the network can be directly reflected in the original image, so that the intuitiveness of the effect is greatly improved.
The above description is only a preferred embodiment of the present invention, and the scope of the present invention is not limited to the above embodiment, but equivalent modifications or changes made by those skilled in the art according to the present disclosure should be included in the scope of the present invention as set forth in the appended claims.
Claims (10)
1. A human body target detection method facing a radar signal heat map is characterized in that: the method comprises the following steps:
step 1: preprocessing data; the data comprises radio frequency data and video data, the radio frequency data is radio frequency signals sent by a multi-antenna array, the signals contact a target in a target area and are reflected back to an original point, vertical and horizontal bidirectional target space information is obtained according to the reflected signals, signal heat map original data is constructed, and then noise reduction, cutting and marking are carried out; the original data is a numerical matrix generated after each frame of radar signal is directly acquired, wherein the 2 nd, 4 th, 6 th and 8 th dimensions represent horizontal direction data, the 1 st, 3 th, 5 th and 7 th dimensions represent vertical direction data, the horizontal direction data, the 3 th, the 5 th and the 7 th dimensions represent vertical direction data, the vertical direction data are respectively taken out and then processed by image Fourier transform to obtain a heat map matrix, and a horizontal heat map data set and a vertical heat map data set are formed; the video data is obtained by synchronously recording a target scene by using a depth camera, and performing target identification storage and human body diagram extraction on a recorded video image by using a Faster R-CNN network to construct a label set for subsequent network training;
step 2: constructing a feature extraction network based on ResNet-34; according to a residual block structure in the ResNet network, constructing a feature extraction network: taking a network structure of ResNet-34 as a main network, selecting a Basic Block type as a residual Block, setting the number set of the residual blocks as [3,4,6,3], wherein the numerical value represents the number of the residual blocks contained in the Layer of each residual network; the network is used for extracting a characteristic matrix for subsequent network processing;
and step 3: reconstructing a characteristic matrix; establishing a subsequent network architecture after the characteristic extraction network, and reconstructing the characteristic matrix through matrix flattening, deformation operation and a full connection layer to perfect the characteristic matrix;
and 4, step 4: fusing horizontal and vertical heat map features; performing fusion operation on the feature matrix obtained in the step; because the heat map is divided into a horizontal data block and a vertical data block, a block diagram data matrix is obtained after the features of the two heat maps are fused; an additional network architecture which is constructed by the step 3 and the step 4 is called a post-network and is used for processing the output of the feature extraction network;
and 5: constructing a target detection frame; and performing image segmentation operation on the original image by using a block diagram data matrix output by a post-network so as to realize the visual effect of human body target detection.
2. The method for detecting human targets oriented to radar signal heat maps of claim 1, wherein: the method comprises the following steps: in the step 1, label data are generated by video data through a Fast R-CNN network and are subjected to batch processing operation, so that the target detection regression frame coordinates of each original image are extracted, and finally, the regression frame coordinate data are converted into a Tensor data type through a DataLoader method of a Pyroch, so that a label set is obtained.
3. The method for detecting human targets oriented to radar signal heat maps of claim 1, wherein: in the step 1, white noise removal is needed to be carried out on the original data of the signal heat map to finish noise reduction processing, the white noise is generally a numerical value of which the data is close to 0 or is lower than the surrounding, and the heat map data is converted into a numerical matrix after the operation is finished; in order to complete the cutting and marking work of the original data, cutting and compressing the numerical matrix generated by each data frame to make the numerical matrix suitable for the network input of the step 2; and the matrix representing the horizontal data is marked and distinguished from the matrix representing the vertical data, respectively. And finally, dividing the processed data into a training set and a test set according to the ratio of 8: 2.
4. The method for detecting human targets oriented to radar signal heat maps of claim 1, wherein: in step 1, for the generated label set, training set and test set, a Pythrch machine learning library is used for data encapsulation, and finally, an encapsulated DataLoader is generated to be loaded in a network.
5. The method for detecting human targets oriented to radar signal heat maps of claim 1, wherein: in the step 2, the feature extraction takes four layers as a main framework and is assisted by an upper pooling Layer and a lower pooling Layer; for network input, firstly, a first convolutional Layer and a pooling Layer are used for scaling a feature map, and then the feature map enters a main unit formed by four layers, each Layer consists of a plurality of residual blocks, and every two convolutional layers in the main unit form a residual block structure in a mapping connection mode; for the network output, processing is performed through a global average pooling layer.
6. The method of detecting human targets oriented to radar signal heat maps of claim 5, wherein: in step 2, the type of the residual Block used by the network is Basic Block, and each residual Block is composed of two 3 × 3 convolutional layers.
7. The method for detecting human targets oriented to radar signal heat maps of claim 1, wherein: in step 3, for the horizontal heat map and the vertical heat map, two feature matrices Out1 and Out2 can be respectively obtained through step 2, and the feature matrices are reconstructed; firstly, performing Flatten operation on multidimensional data, and then performing characteristic processing on a full connection layer FC of 5x4 to achieve matrix optimization; wherein, the parameter of Flatten needs to be changed to adapt to different lengths and widths of network output; after the full connection layer FC, converting the characteristic matrixes of the horizontal heat map and the vertical heat map into final network output; and adopting Reshape operation after the full connection layer for the output display of the test set.
8. The method for detecting human targets oriented to radar signal heat maps of claim 1, wherein: in step 4, the fusion operation comprises two different fusion types of Concat and Add; add fusion operations are used to accomplish horizontal, vertical data fusion.
9. The method for detecting human targets oriented to radar signal heat maps of claim 1, wherein: in step 5, outputting the visual effect according to each frame of image; the background of the image is a two-dimensional plane of an original radar signal range, and then the position of the background of the image is marked according to coordinate set data of each regression frame, so that the effect display of human body target block diagram identification is realized.
10. The method for detecting human targets oriented to radar signal heat maps of claim 1, wherein: in step 5, the two-dimensional plane of the original radar signal range is defined as a null plane in which depth information is cut in an arc-shaped area scanned by the radar signal.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111273690.4A CN113989718A (en) | 2021-10-29 | 2021-10-29 | Human body target detection method facing radar signal heat map |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111273690.4A CN113989718A (en) | 2021-10-29 | 2021-10-29 | Human body target detection method facing radar signal heat map |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113989718A true CN113989718A (en) | 2022-01-28 |
Family
ID=79744525
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111273690.4A Pending CN113989718A (en) | 2021-10-29 | 2021-10-29 | Human body target detection method facing radar signal heat map |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113989718A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115856881A (en) * | 2023-01-12 | 2023-03-28 | 南京邮电大学 | Millimeter wave radar behavior sensing method based on dynamic lightweight network |
| CN117017276A (en) * | 2023-10-08 | 2023-11-10 | 中国科学技术大学 | Real-time human body tight boundary detection method based on millimeter wave radar |
-
2021
- 2021-10-29 CN CN202111273690.4A patent/CN113989718A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115856881A (en) * | 2023-01-12 | 2023-03-28 | 南京邮电大学 | Millimeter wave radar behavior sensing method based on dynamic lightweight network |
| CN117017276A (en) * | 2023-10-08 | 2023-11-10 | 中国科学技术大学 | Real-time human body tight boundary detection method based on millimeter wave radar |
| CN117017276B (en) * | 2023-10-08 | 2024-01-12 | 中国科学技术大学 | Real-time human body tight boundary detection method based on millimeter wave radar |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zamanakos et al. | A comprehensive survey of LIDAR-based 3D object detection methods with deep learning for autonomous driving | |
| CN114202696B (en) | SAR target detection method and device based on context vision and storage medium | |
| CN109685768B (en) | Pulmonary nodule automatic detection method and system based on pulmonary CT sequence | |
| CN108460403A (en) | The object detection method and system of multi-scale feature fusion in a kind of image | |
| CN110688905B (en) | Three-dimensional object detection and tracking method based on key frame | |
| CN111898432B (en) | Pedestrian detection system and method based on improved YOLOv3 algorithm | |
| CN113673510B (en) | Target detection method combining feature point and anchor frame joint prediction and regression | |
| CN111401293B (en) | Gesture recognition method based on Head lightweight Mask scanning R-CNN | |
| CN113989718A (en) | Human body target detection method facing radar signal heat map | |
| CN109977968B (en) | SAR change detection method based on deep learning classification comparison | |
| WO2023030182A1 (en) | Image generation method and apparatus | |
| TWI745204B (en) | High-efficiency LiDAR object detection method based on deep learning | |
| CN113436210B (en) | Road image segmentation method fusing context progressive sampling | |
| CN111160291A (en) | Human eye detection method based on depth information and CNN | |
| CN114495029A (en) | Traffic target detection method and system based on improved YOLOv4 | |
| CN114332098A (en) | Carotid artery unstable plaque segmentation method based on multi-sequence magnetic resonance image | |
| CN112507893A (en) | Distributed unsupervised pedestrian re-identification method based on edge calculation | |
| CN110516527B (en) | Visual SLAM loop detection improvement method based on instance segmentation | |
| CN114332942A (en) | Night infrared pedestrian detection method and system based on improved YOLOv3 | |
| CN114283326A (en) | Underwater target re-identification method combining local perception and high-order feature reconstruction | |
| CN114218999A (en) | Millimeter wave radar target detection method and system based on fusion image characteristics | |
| CN117292324A (en) | Crowd density estimation method and system | |
| CN114077892A (en) | Human body skeleton sequence extraction and training method, device and storage medium | |
| CN114612456B (en) | Billet automatic semantic segmentation recognition method based on deep learning | |
| CN116343041A (en) | Ship classification method based on feature level fusion of HRRP and SAR data |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |