Technology and equipment for searching and positioning wounded based on unmanned aerial vehicle

Abstract

The invention discloses a wounded person searching and positioning technology and equipment based on an unmanned aerial vehicle, and relates to the technical field of artificial intelligence, computer vision and unmanned aerial vehicle emergency rescue. The system is composed of an unmanned aerial vehicle platform, a sensor module, an image processing unit, an artificial intelligence algorithm module and a ground control station. The unmanned aerial vehicle carries a multi-mode sensor to execute autonomous cruise, and disaster area images and environment data are collected in real time; the image processing unit introduces improved deep learning algorithms such as AIA-guide Queries and Enhanced Pixel Decoder to carry out pixel-level segmentation and accurate positioning on the wounded, and synchronously outputs semantic association information of the wounded and dangerous factors such as a fire source and a collapse object; and the ground control station visualizes the result on the GIS platform in real time, and generates an optimal rescue path based on dynamic planning. The method can significantly reduce the leak detection and false detection rate under the background conditions of strong shielding, low illumination and high dynamic, improves the detection robustness and positioning precision in a complex environment, and is suitable for scenes of emergency search and rescue after disasters and the like.

Description

Technical Field

[0001] The present invention relates to the intersection of artificial intelligence, computer vision, and drone emergency rescue technology. The invention specifically discloses a technology and equipment for searching and locating the wounded based on drones, which is used to achieve rapid detection, precise positioning, and environmental risk analysis of the wounded in complex post-disaster scenarios. Background Art

[0002] Natural disasters, major accidents, and emergencies often cause widespread destruction, posing enormous challenges to the search and rescue of the wounded. At the disaster site, adverse factors such as collapsed buildings, blocked roads, and interrupted communications significantly increase the difficulty of rescue operations. Current search and rescue methods rely on manpower and ground equipment, and suffer from problems such as limited coverage, inefficient information acquisition, and insufficient environmental adaptability, making it difficult to meet the needs of rapid response within the golden rescue time. Searching for the wounded is the primary link in the rescue operation, and the difficulty lies in quickly discovering and accurately locating trapped people. However, traditional methods are much less effective when faced with complex terrain, severe weather, and low visibility at night. In addition, existing technical means lack the ability to analyze the semantic association between the wounded and the surrounding environment, which can easily lead to missed or misdetected wounded, delaying rescue efforts.

[0003] In recent years, drones equipped with high-definition RGB cameras and infrared thermal imagers for aerial inspections have become a new trend. Mainstream solutions often combine single-stage bounding box detection networks such as YOLO-v5 / YOLO-v8 or heat source centroid-based localization algorithms to quickly identify suspected casualties. However, because these solutions only output rectangular bounding boxes and rely primarily on single-channel information, they are prone to missed detections in situations with strong occlusion, low contrast, or when the casualty's body is covered. Furthermore, homogeneous backgrounds often lead to false detections. Furthermore, traditional feature pyramids lose small object details during the multi-scale fusion process, failing to provide coherent pixel-level outlines. They also lack explicit modeling of the spatial relationship between casualties and hazardous factors such as fire sources and collapsed objects, resulting in unreliable subsequent environmental assessments and rescue path planning. Therefore, improving the accuracy of casualty detection and segmentation while ensuring real-time performance, while simultaneously outputting semantically related information to hazardous factors, has become a key technical challenge that needs to be addressed in the field of drone search and localization. Summary of the Invention

[0004] To address the shortcomings of existing drone-based casualty search and localization technologies, such as missed detections, high false detection rates, and a lack of pixel-level spatial correlation information under strong occlusion, low light, and high-dynamic background conditions, this paper proposes a drone-based casualty search and localization technology and equipment. The training method introduces semantically related AIA-guided Queries (AIA) query vectors and a multi-scale Enhanced Pixel Decoder (Enhanced Pixel Decoder) feature decoder, enabling the model to explicitly learn the associations between casualties and hazardous factors such as fire sources and collapsed objects during the training phase. Furthermore, the model preserves small target details and corrects for perspective offsets during the multi-scale fusion process. The localization method utilizes the query vectors during the inference phase to amplify potential casualty signals and suppress background noise, while simultaneously outputting a continuous, complete, high-resolution casualty mask and a spatial distribution map of the casualty mask and its relationship to hazardous factors, significantly improving detection robustness and localization accuracy in complex environments.

[0005] Technical problems that need to be solved: (1) Accurately detect and locate the injured under strong occlusion, low light, sloping terrain and high dynamic background conditions; (2) Establish semantic associations between the injured and surrounding dangerous factors such as fire sources and collapsed objects to reduce missed detections and false detections and improve scene discrimination capabilities; (3) Output high-resolution injured person masks and integrate them with environmental analysis results to provide reliable data support for real-time rescue path planning and priority decision-making at the ground control station.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: the technology and equipment for searching and locating the wounded based on drones are composed of a drone platform, a sensor module, an image processing unit, an artificial intelligence algorithm module and a ground control station. These parts work together to complete the detection, positioning and rescue planning of the wounded in complex environments.

[0007] The drone platform is the system's core execution engine, responsible for patrolling and collecting data in post-disaster areas. The drone uses autonomous path planning to cover high-priority areas and possesses flight stability in highly dynamic environments, such as strong winds and confined spaces. Equipped with high-definition cameras and infrared thermal imaging equipment, the platform supports both daytime and nighttime operations, enabling precise target detection in low-light conditions.

[0008] The sensor module achieves comprehensive awareness of the casualty and the environment through multimodal data acquisition. It includes a high-resolution RGB camera, an infrared thermal imager, and environmental detection sensors. The RGB camera captures high-definition images, the thermal imager quickly identifies the target's body temperature distribution, and the environmental sensor assists in analyzing obstacles and hazards, providing environmental information for rescue efforts.

[0009] The image processing unit uses real-time drone image data, combined with preprocessing techniques such as denoising, enhancement, and cropping, to generate high-quality input images. For casualty detection in complex environments, a target segmentation algorithm is employed to improve target detection accuracy through feature extraction and multi-scale information aggregation.

[0010] The AI ​​algorithm module utilizes an improved deep learning algorithm for precise detection and segmentation of casualties. This includes a semantically relevant query mechanism and enhanced pixel decoding technology to improve positioning accuracy and retain segmentation details in complex scenarios. By optimizing the loss function, the system effectively reduces false positives and missed detections, enhancing its ability to model the relationship between casualties and their environment.

[0011] The ground control station serves as the system's command center, responsible for drone mission planning, data processing, and visualization. Using a GIS platform, the control station displays real-time casualty distribution and environmental analysis results. It also uses dynamic programming algorithms to optimize rescue routes, prioritizing high-risk patients.

[0012] The use of drone-based casualty search and location technology and equipment includes the following steps:

[0013] First, the ground control station plans the flight mission, determines the target area and operational requirements in the post-disaster area, and formulates a high-priority cruise path based on the characteristics and environmental conditions of the disaster area.

[0014] The drone then performs autonomous cruising according to the planned mission path, collecting multimodal data of the target area. The drone's onboard high-definition cameras and infrared thermal imaging equipment capture images and environmental data in real time, transmitting this data to the ground control station via wireless communication.

[0015] The image processing unit preprocesses the received multimodal data, including denoising, enhancement, and cropping, to ensure input image quality. Subsequently, multimodal data fusion technology is used to fuse the data from the RGB image and infrared thermal imaging equipment to improve the accuracy of casualty detection.

[0016] The AI ​​algorithm module, based on improved deep learning technology, accurately detects and segments casualties. First, the algorithm uses object segmentation to extract the casualty's precise location and posture information, and then combines semantic association analysis to extract features of their surroundings. This information provides data support for subsequent positioning and environmental analysis.

[0017] The drone platform uses its onboard environmental monitoring sensors to monitor the target area for hazardous conditions, such as fire sources, collapsed buildings, and other obstacles, in real time. The system uses correlation analysis algorithms, combined with casualty location data, to generate a spatial distribution diagram of the casualty and the surrounding environment. It also identifies potential hazards and supports rescue route planning.

[0018] All casualty location information, segmentation results, and environmental analysis data are uploaded to the GIS platform for real-time visualization. Based on a dynamic programming algorithm, the system automatically optimizes rescue routes, prioritizing high-risk casualties. This information is then transmitted to the ground control station in real time, assisting rescuers in developing a scientific, prioritized rescue strategy.

[0019] As a preferred embodiment of the present invention, the specific process of environmental hazard analysis based on a drone platform includes: During navigation, the drone uses environmental detection sensors to acquire information about fire sources, collapsed buildings, and other obstacles. By combining casualty segmentation masks with environmental hazard characteristics, the system uses an association analysis algorithm to generate a spatial distribution relationship map between casualties and hazardous objects. This relationship map is visualized on a GIS platform, intuitively displaying hazardous areas and providing a basis for rescue route planning.

[0020] As a preferred embodiment of the present invention, an improved ResNet-50 network is first used as the backbone during the patient segmentation process to extract features from the input multimodal image and output multi-scale feature maps, including resolutions of C3, C4, and C5. The C4 layer is typically used to generate a "mid-level feature map," which integrates the patient's local details and overall morphology. Based on this mid-level feature map, the algorithm calculates the semantic association between pixels and "patient prior features" and selects the pixel locations with the highest association to generate AIA-guided queries. These queries, along with high-resolution pixel features, are then input into the Dual-path Transformer decoder. This decoder, composed of multiple submodules such as cross-attention, self-attention, and mask attention, progressively filters out background noise through multiple rounds of interaction and outputs a patient segmentation mask. Because the queries focus on the patient region and are deeply coupled with the multi-scale features, this method significantly reduces missed and false detections under conditions of occlusion, low light, or cluttered backgrounds.

[0021] As a preferred embodiment of the present invention, to accurately focus on the injured area under conditions of occlusion or dynamic scene changes, this system introduces a "semantic association analysis between the injured and environmental risk factors" mechanism in the mid-level feature map C4. First, two lightweight auxiliary classification branches are set up in parallel on C4 to generate a pixel-level prior probability map of the injured person (PI) and a probability map of environmental risk factors (PH), respectively. Risk factor categories include open flames, collapsed buildings, and sharp debris. Then, a semantic association is calculated for each pixel i.

[0022]

[0023] Where α is a learnable weight, N(i,r) represents the neighborhood window with radius r centered on i, and w ijis a Gaussian weight based on the distance between pixels. This correlation not only makes full use of the prior information of the injured in the pixel itself, but also compensates by the neighborhood danger probability when the injured are blocked, ensuring that the potential injured area is not missed. i Select N from high to low a Pixel embedding generates AIA-guided Queries and compares them with N b The concatenated learnable query vectors are input into the Dual-path Transformer decoder to achieve joint modeling and fine segmentation of the injured and environmental risk factors, providing highly reliable spatial distribution information of the injured and dangerous objects for subsequent rescue path planning.

[0024] As a preferred embodiment of the present invention, when decoding multi-scale features, the EnhancedPixel Decoder structure is used to fully retain the fine information of small targets and complex scenes. The Feature Reference Module (FRM) performs attention weighting on the horizontal connection features from the feature pyramid network, suppresses irrelevant areas and retains detailed features; the Feature Enhancement Module (FEM) fuses contextual information after introducing multi-scale pooling to enhance the significance of the wounded target in local and global scenes; the Feature Calibration Module (FCM) uses deformable convolution to achieve spatial alignment, corrects the feature offset generated by the drone under overhead or oblique shooting angles, and ensures the accuracy of segmentation in different perspectives. Through the above-mentioned multi-module series connection and hierarchical fusion, the Enhanced Pixel Decoder can significantly reduce the loss of small target details while retaining high semantic features, and is particularly effective for detecting wounded targets under partial occlusion.

[0025] As a preferred embodiment of the present invention, a dual-path Transformer decoder is introduced to achieve top-down semantic reasoning and bottom-up mask refinement during the segmentation phase. The decoder maintains two information streams simultaneously at each layer: the query path stores instance query vectors for global semantic aggregation and category determination; and the pixel path retains pixel embeddings from high-resolution feature maps for fine-grained boundary characterization. Each layer first performs cross-attention, where the query reads local context from the pixel and updates the query semantics. It then performs self-attention to achieve instance de-redundancy and global coordination between queries. Finally, mask attention is performed, using the coarse mask generated by the current query as a weight to guide the pixel path to only update pixel features related to the target, thereby suppressing background noise and refining edges. The mask prediction output by the previous layer is explicitly fed into the cross-attention layer of the next layer, achieving iterative convergence from coarse to fine. At the same time, each layer only uses a lightweight feedforward network and layer normalization to ensure that the decoding depth is increased while maintaining real-time inference speed. After multiple layers of loops, the Dual-path Transformer simultaneously outputs the category score and high-resolution segmentation mask for each query. During the training phase, the correspondence between the prediction and the true label is established, and end-to-end optimization is performed using the combined category cross entropy, Dice Loss, and associated IoU loss, thereby significantly improving the accuracy and robustness of casualty detection under occlusion, low light, and cluttered background conditions.

[0026] As a preferred embodiment of the present invention, the patient location and classification process based on multimodal data fusion specifically includes the following: the system inputs RGB images and thermal imaging data collected by drones into a multimodal fusion module. RGB features are used to accurately locate the patient's outline, while infrared thermal imaging features are used to enhance the patient's heat source area. A self-attention mechanism learns the complementary characteristics of the two modalities, and the resulting multimodal features are ultimately fused for patient classification and location, effectively improving detection accuracy in complex environments.

[0027] The present invention has the following beneficial effects:

[0028] 1. By using drone platforms to conduct rapid aerial inspections of disaster areas, combined with the target detection model obtained by the training method of the present invention, high-resolution masks and coordinate information of the injured can be accurately output over a large area and in a short period of time, providing strong technical support for rapid response within the golden rescue time.

[0029] 2. By introducing AIA-guided Queries and Enhanced Pixel Decoder, the present invention explicitly learns the semantic association between casualties and risk factors during the model training phase, and retains small target details and corrects perspective offsets during the inference phase, significantly reducing missed detection and false detection rates in complex backgrounds and improving detection robustness.

[0030] 3. The system visualizes real-time positioning results and environmental risk analysis on the ground control station GIS platform, and combines dynamic programming algorithms to generate the optimal rescue path, allowing commanders to intuitively obtain information on the distribution of casualties and risk levels, thereby making scientific and efficient rescue decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0032] Figure 1 This is a schematic diagram of the process of searching and locating injured persons based on drones in the present invention.

[0033] Figure 2 It is the casualty segmentation process based on AIA-guided Queries.

[0034] Figure 3 It is the network architecture of the Dual-path Transformer decoder.

[0035] Figure 4 It is the network architecture of the key modules in Enhanced Pixel Decoder.

[0036] Figure 5 It is a multi-scale feature decoding process based on Enhanced Pixel Decoder. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by ordinary technicians in this field without making creative efforts are all within the scope of protection of the present invention.

[0038] like Figure 1As shown, the technology and equipment for searching and locating the wounded based on drones involved in the present invention are composed of a drone platform, a sensor module, an image processing unit, an artificial intelligence algorithm module and a ground control station. These parts work together to complete the detection, positioning and rescue planning of the wounded in complex environments. The drone platform is the execution core of the system and is responsible for cruising and data collection in the post-disaster area. The platform adopts a quadcopter structure, and the main body of the fuselage is made of lightweight carbon fiber material, which not only ensures the impact resistance of the body but also reduces the overall weight to adapt to the flight requirements in complex environments. The drone supports autonomous path planning, can cover high-priority areas in post-disaster environments, and dynamically adjust based on real-time environmental data. To adapt to operations at night and in low-light conditions, the platform is equipped with a high-brightness LED light group with a light coverage range of up to 30 meters and can automatically adjust the light intensity according to mission requirements.

[0039] The sensor module of the present invention adopts a modular and expandable design, which can quickly replace or expand the equipment required for different tasks. The sensor module of this embodiment includes a high-resolution RGB camera, an infrared thermal imager, an environmental detection sensor, etc. The RGB camera is used to capture high-definition images, supports multiple optical zoom, and can achieve high-precision visual perception in complex environments. The infrared thermal imager can quickly identify the body temperature distribution of the wounded, and combine the RGB image for multimodal fusion to enhance the accuracy of target detection. The environmental detection sensor is used to monitor dangerous environmental information such as obstacles, fire sources, collapsed buildings, etc. in the disaster area in real time, and provide data support for planning safe rescue routes. All sensors are connected to the drone platform through a universal interface to ensure stable power supply and data transmission.

[0040] The image processing unit (IPU) is the data processing core of the entire system. It utilizes a high-performance embedded computing device equipped with a GPU acceleration module, specifically designed to process the massive amounts of sensor data collected by drones. During data preprocessing, the system performs denoising, enhancement, cropping, and multimodal fusion on the collected image, audio, and infrared data to ensure the quality of the generated input images. For casualty detection in complex environments, an object segmentation algorithm is employed to improve detection accuracy through feature extraction and multi-scale information aggregation. The IPU's intelligent processing capabilities enable the entire system to adapt to various post-disaster environments, performing real-time data analysis and providing feedback to the ground control station.

[0041] As shown in Figure 2, the AI ​​algorithm first uses a modified ResNet-50 backbone to extract feature maps at multiple scales. It then generates AIA-guided queries based on pixel semantic associations in mid-level feature maps, explicitly focusing on the injured area and rejecting background noise. These queries, along with the multi-scale features, are fed into the Dual-path Transformer decoder, shown in Figure 3. The query path aggregates global semantics and determines the category, while the pixel path maintains high-resolution embeddings to refine edges. Both paths converge layer by layer through the iterative effects of cross-attention, self-attention, and mask attention, outputting a coherent and complete injured person mask and category score. The mask features then pass through the Enhanced Pixel Decoder, shown in Figures 4 and 5. This decoder uses FRM to suppress irrelevant regions, FEM to introduce multi-scale pooling to supplement context, and FCM to employ deformable convolutions for spatial alignment. This approach preserves high-level semantic information while significantly reducing the loss of detail in small objects and correcting for feature shifts caused by overhead or oblique drone photography. After segmentation is completed, the algorithm spatially superimposes the injured person's mask with dangerous features such as fire sources and collapsed objects, uses a graph convolutional network to establish an adjacency matrix between the injured and dangerous factors, and generates a relationship diagram, which is uploaded to the ground control station GIS platform in real time.

[0042] The ground control station is the command center for the entire casualty search and location system, responsible for drone mission planning, data processing, and visualization analysis of results. The control station uses a GIS platform to display the distribution of casualties, environmental analysis results, and optimal rescue routes in real time. It also optimizes rescue operations based on a dynamic programming algorithm to ensure that high-risk casualties are rescued quickly and effectively. The ground control station maintains real-time communication with the drone via a portable terminal. Users can set mission modes and fine-tune flight paths on the terminal interface to adapt to the complex and changing post-disaster environment. During mission execution, the control station can monitor all sensor data and analysis results in real time, generating auxiliary decision-making tools such as casualty distribution heat maps and injury priority classification tables to help rescue personnel make scientific rescue decisions.

[0043] The embodiments of the present invention disclosed above are intended only to illustrate the present invention. They do not describe all details in detail, nor do they limit the present invention to specific embodiments. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A technology and equipment for searching and locating casualties based on drones, characterized in that: include: The drone platform is used for autonomous patrol in post-disaster areas, multimodal data collection and wireless transmission; The sensor module is mounted on the UAV platform and includes a high-resolution RGB camera, an infrared thermal imager, and an environmental detection sensor to collect multimodal data from the disaster area. The image processing unit performs denoising, enhancement, cropping and multimodal fusion preprocessing on the collected image data to form high-quality image input; The artificial intelligence algorithm module uses an improved deep learning method, a query mechanism based on semantic association, and enhanced pixel decoding technology to accurately detect and segment the wounded, generating a spatial distribution relationship map between the wounded and the environment. The ground control station is used for drone mission planning, real-time data processing and result visualization. It displays the distribution of casualties and environmental analysis results through the GIS platform and dynamically optimizes the rescue route.

2. The UAV-based casualty search and location technology and equipment according to claim 1 is characterized in that: The drone platform has a quadcopter structure and a body made of lightweight carbon fiber material. It supports high-dynamic flight and autonomous path planning, and integrates a high-definition RGB camera, an infrared thermal imager, and a dimmable high-brightness LED light group to adapt to night and low-light operations.

3. The UAV-based casualty search and location technology and equipment according to claim 1 is characterized in that: The sensor module adopts a modular and scalable design. All sensors achieve stable power supply and data transmission through a universal interface. The RGB camera has a multi-fold optical zoom function, the infrared thermal imager is used to identify the temperature distribution of the injured in real time, and the environmental detection sensor is used to sense fire sources, obstacles and collapsed buildings.

4. The UAV-based casualty search and location technology and equipment according to claim 1, characterized in that: The image processing unit is a high-performance embedded computing device equipped with a GPU acceleration module, which improves the accuracy of casualty detection and segmentation through feature extraction, multi-scale information aggregation and multimodal fusion technology.

5. The UAV-based casualty search and location technology and equipment according to claim 1 is characterized in that: The casualty segmentation method of the artificial intelligence algorithm module includes: An improved ResNet-50 network is used to extract multi-scale features from multimodal input images; Generate AIA-guided Queries based on pixel semantic association on the middle-level feature map; The queries and multi-scale features are input into the Dual-path Transformer decoder for multiple rounds of interaction to output a high-resolution segmentation mask of the injured person.

6. The UAV-based casualty search and location technology and equipment according to claim 1 is characterized in that: The environmental hazard analysis method of the artificial intelligence algorithm module is specifically as follows: The drone collects environmental hazard information in real time during the cruise process; The casualty segmentation mask is integrated with the environmental hazard characteristics, and the spatial distribution relationship diagram of casualties and hazard factors is generated using the association analysis algorithm. The relationship diagram is visualized on the GIS platform to provide a basis for rescue route planning.

7. The UAV-based casualty search and location technology and equipment according to claim 1 is characterized in that: The Enhanced Pixel Decoder multi-scale feature decoding method includes: Generate reference feature maps through lateral connections of the feature pyramid network; Use the attention mechanism of Feature Reference Module (FRM) to highlight the details of the injured area; Introducing multi-scale pooling and contextual information into the Feature Enhancement Module (FEM) to enhance the ability to capture small objects; Deformable convolution is used for spatial alignment via Feature Calibration Module (FCM) to improve segmentation mask accuracy.

8. The UAV-based casualty search and location technology and equipment according to claim 1 is characterized in that: The multimodal data fusion wounded positioning method includes: The drone simultaneously collects RGB images and infrared thermal imaging data and inputs them into the multimodal fusion module; RGB features are used to accurately depict the outline of the casualty, and infrared features are used to enhance the casualty's heat source area; The complementarity of the two modal features is learned through the self-attention mechanism to generate fusion features for casualty classification and positioning.

9. A working method using the UAV-based casualty search and location technology and equipment according to claim 1, characterized in that: The following steps are involved: The ground control station plans the flight mission, determines the target area and cruise path; The UAV cruises autonomously and collects multimodal image data in real time, transmitting it to the ground control station; The image processing unit preprocesses the data and fuses the multimodal data; The artificial intelligence algorithm module accurately detects casualties and generates a spatial distribution relationship map; The GIS platform visualizes the location of the injured, environmental analysis results in real time, and optimizes rescue routes; The ground control station dynamically plans rescue strategies based on real-time information, giving priority to rescuing high-risk casualties.

10. The method of use according to claim 9, characterized in that: During the cruise, the drone continuously monitors dangerous environmental factors such as fire sources, obstacles and collapsed buildings, and correlates and analyzes such information with real-time updated location information of the injured to further optimize the rescue route and action plan.

Images