Target detection method and apparatus

Abstract

The application provides a target detection method and device, and relates to the technical fields of computer vision and remote sensing detection. The method comprises the following steps: generating a first probability heat map reflecting a parking rule of a target based on historical observation image data of a region of interest; generating a second probability heat map reflecting an environmental action relationship based on the historical observation image data and geographic information system data of the region of interest; and detecting the region of interest by using a trained target detection model with the first probability heat map and the second probability heat map as weight factors to obtain a target detection result.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and remote sensing technology, and particularly relates to a target detection method and apparatus. Background Technology

[0002] The management of specific types of mobile targets (such as tank trucks, hazardous materials transport vehicles, and container trucks) is crucial for regional security, logistics scheduling, and emergency response. The demand for accurate detection, tracking, and behavioral analysis of these targets is also gradually increasing. Existing target detection technologies are mainly divided into two categories: methods based on general deep learning models and methods based on scene modeling and target tracking.

[0003] Common methods based on general-purpose deep learning models include Faster R-CNN, YOLO, and SSD. These methods, trained on large general-purpose datasets such as the Common Objects in Context (COCO), can effectively detect specific types of moving objects in images. However, these methods are passive detection methods, and their performance heavily depends on the quantity and quality of the training data. In complex real-world scenarios, they are easily affected by factors such as changes in lighting, weather, and object occlusion, leading to both false positives and false negatives. The performance of these methods deteriorates sharply, especially in tasks involving small objects that are far away and have a small pixel footprint.

[0004] Common scene-based modeling and target tracking methods typically use Gaussian mixture models for background modeling, combined with Kalman filtering or correlation filtering for target tracking. These methods perform well in handling moving targets, but may confuse long-term stationary or parked targets with the background, resulting in missed detections or tracking failures. Furthermore, these methods cannot distinguish the specific type of target, making it difficult to perform targeted detection of specific targets such as tanker trucks. Summary of the Invention

[0005] In view of this, the present invention provides a target detection method and apparatus, which solves the technical problems existing in the target detection methods in dealing with low imaging quality, complex operating scenarios, and static target detection tasks.

[0006] The first aspect of the present invention provides a target detection method, comprising: generating a first probabilistic heatmap reflecting the parking pattern of a target based on historical observation image data of a region of interest; generating a second probabilistic heatmap reflecting the environmental interaction based on historical observation image data and geographic information system data of the region of interest; and using the first probabilistic heatmap and the second probabilistic heatmap as weighting factors, employing a trained target detection model to detect the region of interest, thereby obtaining a target detection result.

[0007] According to an embodiment of the present invention, generating a first probability heatmap reflecting the parking pattern of targets based on historical observation image data of the region of interest includes: dividing the region of interest into multiple grid cells, and determining the initial parking probability of the target in each grid cell according to the historical observation image data; correcting the initial parking probability according to the spatial distance between each grid cell and the historical target location to obtain the corrected parking probability of each grid cell; and normalizing the corrected parking probability and mapping it to the image space to obtain the first probability heatmap.

[0008] According to an embodiment of the present invention, determining the initial parking probability of a target in each grid cell based on historical observation image data includes: determining a weight for each historical observation frame, the weight being negatively correlated with the time difference between the historical observation frame and the last frame; and using the weight to perform a weighted summation of the occurrence of a target in each grid cell in each historical observation frame to obtain the initial parking probability of the grid cell.

[0009] According to an embodiment of the present invention, correcting the initial parking probability based on the spatial distance between each grid cell and the historical target location includes: determining the minimum spatial distance between each grid cell and all historical target locations; generating an exponential decay coefficient based on the minimum spatial distance, wherein the exponential decay coefficient decreases as the minimum spatial distance increases; and correcting the initial parking probability using the exponential decay coefficient to obtain the corrected parking probability.

[0010] According to an embodiment of the present invention, generating a second probabilistic heatmap reflecting environmental interactions based on historical observation image data and geographic information system (GIS) data of the area of ​​interest includes: dividing the area of ​​interest into multiple grid cells and extracting points of interest (POIs) from the GIS data; using a large language model to perform semantic association reasoning on the attributes of the POIs, parking behavior data extracted from the historical observation image data, and dynamic environmental parameters to obtain the association weight, distance decay function, and environmental correction coefficient between each grid cell and each POI; determining the initial probability correction value for each grid cell based on the association weight, distance decay function, and environmental correction coefficient; and normalizing the initial probability correction value and mapping it to the image space to obtain the second probabilistic heatmap.

[0011] According to an embodiment of the present invention, semantic association reasoning using a large language model includes: constructing structured prompt words from the attributes of interest points, parking behavior data, and dynamic environment parameters, and inputting them into the large language model; outputting association rules in natural language form through the large language model, the association rules including the basic weights of each interest point, distance decay conditions, filtering rules for abnormal parking records, and correction rules for dynamic environment parameters; and converting the association rules into association weights, distance decay functions, and environment correction coefficients between each grid cell and each interest point.

[0012] According to an embodiment of the present invention, determining the initial probability correction value of each grid cell includes: for each point of interest, multiplying the association weight by the distance decay function to obtain the contribution component of the point of interest to the grid cell; summing the contribution components of all points of interest and multiplying them by the environment correction coefficient to obtain the initial probability correction value of the grid cell.

[0013] According to an embodiment of the present invention, using a first probability heatmap and a second probability heatmap as weighting factors, and employing a trained target detection model to detect the region of interest includes: using the target detection model to detect the region of interest, obtaining the target detection box and initial confidence level; determining the center point of the detection box, and obtaining the first probability value and the second probability value corresponding to the center point from the first probability heatmap and the second probability heatmap, respectively; performing a weighted correction on the initial confidence level based on the first probability value and the second probability value to obtain the final confidence level; assigning the final confidence level to the detection box and outputting it as the target detection result.

[0014] According to an embodiment of the present invention, the method further includes: adding the confirmed target detection result to the historical observation image data, and updating the first probability heatmap and the second probability heatmap.

[0015] A second aspect of the present invention provides a target detection device, comprising: a first generation module, configured to generate a first probabilistic heatmap reflecting target parking patterns based on historical observation image data of a region of interest; a second generation module, configured to generate a second probabilistic heatmap reflecting environmental interactions based on historical observation image data and geographic information system data of the region of interest; and a target detection module, configured to use the first and second probabilistic heatmaps as weighting factors and a trained target detection model to detect the region of interest, thereby obtaining target detection results.

[0016] According to the target detection method of the present invention, a prior heat map (i.e., a first probability heat map) reflecting the parking pattern of the target and a heat map of the relationship with the environment (i.e., a second probability heat map) are constructed based on data obtained from long-term observation. The two types of heat maps are used as weighting factors and integrated with the back-end inference decision-making process of the real-time detection system, which can achieve optimized allocation of detection resources and accurate discrimination of detection results. Attached Figure Description

[0017] The above and other objects, features and advantages of the present invention will become more apparent from the following description of embodiments of the invention with reference to the accompanying drawings, in which:

[0018] Figure 1 A flowchart illustrating a target detection method according to an embodiment of the present invention is shown schematically;

[0019] Figure 2A flowchart illustrating a target detection method according to yet another embodiment of the present invention is shown.

[0020] Figure 3 A first probability heatmap generated according to an embodiment of the present invention is illustrated schematically;

[0021] Figure 4 A second probability heatmap generated according to an embodiment of the present invention is illustrated schematically;

[0022] Figure 5 A schematic diagram illustrating the structure of a target detection device according to an embodiment of the present invention is shown.

[0023] Figure 6 A block diagram of an electronic device suitable for implementing a target detection method according to an embodiment of the present invention is shown schematically. Detailed Implementation

[0024] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the invention for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0025] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0026] Existing object detection technologies are mainly divided into two categories: methods based on general deep learning models and methods based on scene modeling and object tracking. The drawbacks of existing technologies are:

[0027] (1) Poor detection performance under low imaging quality: Existing methods rely only on information from single-frame or sequential image data during processing, ignoring the regularity of the existence of specific targets in the scene. Under conditions such as low image capture frequency, low resolution, and poor image quality, it is impossible to obtain effective feature information for detection.

[0028] (2) Poor robustness in complex scenarios: When the appearance features of the target are not obvious, the method based on the general deep learning model is prone to obvious missed detection. In complex operating scenarios, the target may have different appearance features than in the normal scenario due to factors such as lighting, precipitation, occlusion, etc., thereby reducing the detection rate.

[0029] (3) Weak ability to handle stationary targets: The method based on scene modeling and target tracking may confuse stationary targets with the background, resulting in the loss of stationary targets.

[0030] To address the shortcomings of existing technologies, this invention aims to propose a precise target detection method based on the relationship between target parking patterns and environmental interactions. This method integrates the long-term behavioral patterns and spatial distribution prior knowledge of specific targets into the real-time detection process, thereby significantly improving the detection performance and robustness for such targets in complex scenarios. It solves the technical challenges in dealing with low-image-quality, complex operating scenarios, and stationary target detection tasks. The target detection method will be described in detail below with reference to embodiments and accompanying drawings.

[0031] Figure 1 A flowchart illustrating a target detection method according to an embodiment of the present invention is shown schematically. Figure 1 As shown, the method may include operations S110 to S130.

[0032] In operation S110, a first probability heatmap reflecting the parking pattern of the target is generated based on historical observation image data of the area of ​​interest.

[0033] In embodiments of the present invention, the area of ​​interest refers to the spatial range in which the target may appear. Typical scenarios include chemical industrial parks, port terminals, logistics hubs, main roads for the transportation of hazardous chemicals, and surrounding parking areas. The target may be a mobile target with specific management requirements, such as a tanker truck, a hazardous chemical transport vehicle, or a container truck.

[0034] Historical observation image data can be obtained by using satellite remote sensing, drone inspections, or fixed surveillance cameras to collect image sequences of areas of interest over long periods of time and periodically. Each image corresponds to an observation time (frame) and is georegistered to ensure that the image pixel coordinates correspond one-to-one with the real-world geographic coordinates.

[0035] Target placement patterns refer to the probability distribution characteristics of a target's spatial location over a historical observation period. These patterns not only include the target's long-term, frequently appearing habitable areas but also implicitly contain the neighbor diffusion characteristics of random shifts within these habitable areas. By quantifying these patterns, a spatial probability map can be obtained to guide subsequent real-time target detection.

[0036] The first probability heatmap is a heatmap based on target parking patterns. Its core idea is that target parking behavior in a fixed scene exhibits spatiotemporal repetition and local clustering. By analyzing historical observation data and statistically analyzing the frequency of target occurrences at various spatial locations, a prior probability distribution can be constructed. This distribution describes the knowledge of "where the target is most likely to appear" before observation. The first probability heatmap's advantages are: for targets located in high-probability areas but with low initial detection confidence, prior probability can enhance their confidence and reduce missed detections; for false alarm targets located in low-probability areas, its confidence can be relatively weakened, effectively suppressing false detections; furthermore, spatial diffusion processing ensures that even slight shifts in target location can still be covered by the heatmap, reducing reliance on precise positioning.

[0037] In operation S120, a second probabilistic heatmap reflecting environmental interactions is generated based on historical observation image data and geographic information system data of the area of ​​interest.

[0038] In embodiments of the present invention, the geographic information system data of the area of ​​interest refers to geographic information system (GIS) map data containing point of interest (POI) information. POIs include facilities or areas closely related to the target parking behavior, such as oil storage tanks, loading and unloading equipment, chemical plants, dedicated parking spaces, and roads. These POIs typically have information such as functional attributes, spatial extent, and operational constraints.

[0039] Environmental interaction refers to the correlation between target parking behavior and surrounding geographical environmental factors. For example, tanker trucks tend to park in loading and unloading areas near oil storage tanks, and prefer temporary parking areas with rain protection facilities on rainy days. These relationships reflect the influence of the environment on target behavior.

[0040] The second probability heatmap is a probability distribution map generated based on environmental interactions. Its core idea is that target parking behavior is not random but significantly influenced by surrounding geographical environmental factors. The benefits of the second probability heatmap are: for targets located in areas with high environmental relevance, even with low initial detection confidence, prior environmental knowledge can be used to enhance their confidence; conversely, for false alarm targets located in areas with low environmental relevance, their confidence is relatively weakened. Furthermore, the second probability heatmap can dynamically integrate real-time environmental parameters, adapt to scene changes, and improve detection robustness in complex environments.

[0041] In operation S130, the first probability heatmap and the second probability heatmap are used as weighting factors, and the trained target detection model is used to detect the region of interest to obtain the target detection result.

[0042] In embodiments of the present invention, the trained object detection model can be a deep learning-based object detection network (or model) capable of extracting visual features from a single frame image and outputting the bounding box of the target and its initial confidence score. Embodiments of the present invention do not limit the specific type of object detection network (or model). However, in complex scenes, relying solely on single-frame visual features can easily lead to missed or false detections. Therefore, embodiments of the present invention incorporate a first probability heatmap and a second probability heatmap as weighting factors into the model's object detection. That is, the first probability heatmap (reflecting historical parking patterns) and the second probability heatmap (reflecting environmental interactions) are used as prior knowledge to weight and adjust the initial confidence score output by the detection model. The adjusted confidence score better reflects the probability that the target actually exists in the current scene.

[0043] This fusion mechanism effectively compensates for the shortcomings of pure visual detection. On the one hand, the historical parking pattern heatmap utilizes the spatiotemporal repetition of targets, significantly improving the recall rate of targets in high-probability areas. On the other hand, the environmental interaction heatmap utilizes prior geographical knowledge, effectively filtering out false alarms that do not conform to environmental associations. The combined effect of these two mechanisms improves both the accuracy and recall rate of target detection, making it particularly suitable for complex scenarios with clear geographical structures and parking patterns, such as chemical industrial parks, ports, and logistics hubs.

[0044] In an exemplary embodiment, various methods such as weighted product, weighted summation, or adaptive fusion function can be used to combine the probability values ​​of the two heatmaps with the initial confidence level. Specific fusion strategies will be described in detail in subsequent embodiments. The final output target detection result includes the target's location, category, and fused confidence level, which can be further used for downstream tasks such as tracking, behavior analysis, or security management.

[0045] Please see Figure 2 The diagram schematically illustrates a flowchart of a target detection method according to another embodiment of the present invention. The target detection method of this embodiment is... Figure 1 Further refinement and expansion of the target detection method shown.

[0046] like Figure 2 As shown, before generating the first probability heatmap (i.e., a heatmap based on target parking patterns) and the second probability heatmap (i.e., a heatmap based on environmental interactions), target annotation is performed on historical observation image data. Existing target detection models can be used in conjunction with manual review to identify and record the bounding box coordinates and corresponding timestamps of targets (such as tankers) in each frame.

[0047] In the target detection method of the present invention, the generation process of the first probability heatmap may include spatial gridding, parking pattern statistics, normalization and heatmap generation steps.

[0048] First, spatial gridding. For example, for a geo-registered remote sensing image, the area of ​​interest is divided into M×N uniform grid units, where M and N are positive integers. The specific values ​​are set according to the actual scene resolution requirements, and this invention does not limit them.

[0049] Next, the parking pattern is statistically analyzed. First, the initial parking probability of the target in each grid cell can be determined based on historical observation image data. Then, the initial parking probability is corrected to obtain the corrected parking probability for each grid cell.

[0050] In a more conventional implementation, all historical annotation data can be traversed, and the number of frames C(i,j) in which the target appears in each grid cell (i,j) can be counted. Then, the parking probability coefficient P1(i,j) of that grid cell can be calculated as follows:

[0051]

[0052] Where T is the total number of frames of historical data.

[0053] In some embodiments, a time decay factor can be introduced to calculate the initial parking probability, giving higher weight to recent data. Specifically, a weight negatively correlated with the time difference between the historical observation frame and the last frame can be determined for each historical observation frame. Then, the occurrence of the target in the grid cell in each historical observation frame is weighted and accumulated using the weights to obtain the initial parking probability of the grid cell. The calculation formula for the parking probability coefficient in embodiments of the present invention can be expressed as:

[0054]

[0055] Where τ is a time dimension coefficient with a value of 1, and σ k (i,j) represents a function indicating the existence of the target at position (i,j) in the k-th frame. If the target exists, σ k (i,j)=1, otherwise σ k (i,j)=0;ΔT k is the time difference between the k-th frame and the last frame.

[0056] The initial parking probability calculated in any of the above embodiments can be corrected. Specifically, this process may include: first, determining the minimum spatial distance between each grid cell and all historical target locations; then, generating an exponential decay coefficient based on the minimum spatial distance, where the exponential decay coefficient decreases as the minimum spatial distance increases; and finally, correcting the initial parking probability using the exponential decay coefficient to obtain the corrected parking probability. As an example, the corrected parking probability can be expressed as:

[0057]

[0058] Among them, P 1,d (i,j) represents the corrected parking probability, d min Let (i,j) be the Euclidean distance between the position (i,j) and the nearest historical target position. This is the exponential decay coefficient generated based on the minimum spatial distance. By correcting the initial parking probability and assigning a certain probability factor to the area near the target's common locations, the robustness of the detection system can be enhanced.

[0059] Finally, normalization and heatmap generation are performed. For example, the corrected parking probability values ​​of all grid cells are linearly normalized to the [0,1] interval to obtain the normalized probability P. A (i,j). Then, a probability heatmap with the same dimensions as the original scene image is generated. The grayscale or color of each pixel in the heatmap is mapped from the normalized probability value of its corresponding grid. The brighter the region in the heatmap, the greater the probability of the target appearing historically.

[0060] Figure 3 The diagram schematically illustrates a first probability heatmap (presented in grayscale) generated according to an embodiment of the present invention. This heatmap reflects the target's parking pattern, that is, the probability distribution of the target's appearance at various locations in space during historical observation periods. Higher brightness indicates a greater probability of the target (e.g., a tanker truck) appearing in that area, and vice versa. Figure 3 As shown, the original remote sensing image clearly displays multiple large storage tanks and their surrounding facilities. The first probability heatmap uses grayscale encoding, obtaining high probability values ​​(displayed as bright areas) at common target parking locations based on historical data accumulation. Simultaneously, a certain level of probability values ​​also exists in the adjacent areas surrounding common parking locations, thus forming a heatmap reflecting the spatial distribution pattern of target parking. Please continue reading. Figure 2 In the target detection method of the embodiments of the present invention, the generation process of the second probability heatmap may include geographic information system fusion, environmental correlation analysis based on a large model, normalization and heatmap generation steps.

[0061] First, Geographic Information System (GIS) fusion is performed. In this step, the area of ​​interest can be divided into multiple grid cells, and points of interest (POIs) can be extracted from GIS data. The grid cells can be spatially gridded according to the rules in the heatmap generation method based on target parking patterns, which will not be elaborated here. Then, the GIS map data of the scene is acquired, and the locations of key POIs are extracted. Taking a scene with tanker trucks as the target as an example, key POIs could include oil storage tanks, loading and unloading equipment, chemical plant buildings, dedicated parking spaces, roads, etc.

[0062] Next, environmental correlation analysis is performed based on a large model. Specifically, this analysis includes: using a Large Language Model (LLM) to perform semantic correlation reasoning on the attributes of points of interest, parking behavior data extracted from historical observation image data, and dynamic environmental parameters to obtain the correlation weight, distance decay function, and environmental correction coefficient between each grid cell and each point of interest; and determining the initial probability correction value for each grid cell based on the correlation weight, distance decay function, and environmental correction coefficient. The core of this step lies in leveraging the semantic understanding and reasoning capabilities of LLM to automatically mine implicit correlation rules between environmental elements and target parking behavior from multi-source heterogeneous data, thereby generating quantitative parameters for heatmap calculation.

[0063] The following explains the specific process of semantic association reasoning using a large language model.

[0064] First, the attributes of points of interest, parking behavior data, and dynamic environment parameters are constructed into structured prompts and input into the large language model. Attributes, parking behavior data, and dynamic environment parameters constitute multi-source raw data. Transforming multi-source raw data into structured prompts that can be parsed by the LLM ensures the integrity and semantic accuracy of the input information.

[0065] By constructing cue words from the attribute data of Points of Interest (POIs), attribute completion and semantic annotation can be performed on the original GIS map, clarifying key information such as the functional characteristics, security level, and usage constraints of the POIs. Taking an oil tanker truck as an example, an example of constructing cue words from the attribute data of POIs is shown in Table 1 below:

[0066] Table 1

[0067]

[0068] In some embodiments, this step can automatically complete the missing POI attributes in manual annotations using the entity recognition and attribute completion capabilities of LLM. For example, based on the scenario of "port hazardous chemical area", LLM can infer that unlabeled oil storage tanks must meet the requirement of "explosion protection level ≥ Ex d IIB".

[0069] Using historical parking behavior data to construct prompts integrates the spatiotemporal information, behavioral purpose, and environmental variables of the target's historical parking records, avoiding the limitation of traditional data that only contains coordinates and time. Taking tanker trucks as an example, an example of constructing prompts from historical parking behavior data of points of interest is shown in Table 2 below:

[0070] Table 2

[0071]

[0072] Using dynamic environmental parameters as prompts can clearly define the scene boundaries and real-time environmental conditions of the current heatmap generation, providing a scene benchmark for LLM correlation analysis. Taking a tanker truck as an example, an example format for using the dynamic environmental parameters of points of interest as prompts is shown in Table 3 below:

[0073] Table 3

[0074]

[0075] Next, the association rules in natural language form are output through the large language model. The association rules include the basic weights of each interest point, distance decay conditions, filtering rules for abnormal parking records, and correction rules for dynamic environmental parameters.

[0076] The embodiments of this invention employ a domain-fine-tuned large language model, fine-tuned using tens of thousands of labeled data points of the target parking scene, to ensure the model's accuracy in understanding the semantics of the target domain, and perform the following core inference tasks:

[0077] First, LLM, based on the input structured data, outputs association rules with quantified parameters, clarifying the influence weights of different POI attributes on parking probability and the distance decay law. Taking tanker trucks as an example, the output rules can be shown in Table 4 below:

[0078] Table 4

[0079]

[0080] Next, LLM uses semantic judgment to remove noisy records from historical data and reduces the weight of such records on the association model. As an example, the filtering rules and weight adjustment logic can be as follows:

[0081] Filtering rules: If a historical parking record meets the following conditions, such as "stay time < 15 min", "behavioral purpose ≠ operation / waiting for operation", and "no hazardous chemical related POIs (oil storage tanks / loading and unloading equipment, etc.) within 50m", it is judged as an abnormal parking record (such as temporary breakdown parking, driver rest, and other non-target behaviors); Weight adjustment: The association calculation weight of normal parking records is 1.0, and the weight of abnormal parking records is reduced to 0.05 to avoid them interfering with the establishment of the normal association model.

[0082] Next, the LLM dynamically adjusts the association weights of POIs based on real-time changes in environmental parameters and scene constraints. Taking a tanker truck as an example, the example output rules are shown in Table 5 below:

[0083] Table 5

[0084]

[0085] Next, the association rules are converted into association weights, distance decay functions, and environmental correction coefficients between each grid cell and each point of interest.

[0086] The natural language association rules and weight parameters output by LLM are transformed into structured data that can be directly used for heatmap calculations. For example, the association weight W between each grid (i,j) and its surrounding POIs is output. ij (k) (k is the POI identifier). As an example, the matrix format can be as shown in Table 6 below (grid size set to 10m × 10m):

[0087] Table 6

[0088]

[0089] Finally, the initial probability correction value for each grid cell is determined based on the association weight, distance decay function, and environmental correction coefficient.

[0090] Specifically, for each point of interest, the association weight can be multiplied by the distance decay function to obtain the contribution component of the point of interest to the grid cell. Then, the contribution components of all points of interest are summed and multiplied by the environment correction coefficient to obtain the initial probability correction value of the grid cell.

[0091] As an example, the formula for calculating the initial probability correction value is as follows:

[0092]

[0093] Among them, P LLM (i,j) represents the initial probability correction value; W ij (k) represents the association weight between grid (i,j) and POI-k (output by LLM); D ij (k) represents the actual spatial distance from grid (i,j) to POI-k; f(D ij (k) is the distance decay function (based on the rules generated by LLM, such as "f(D)=1 when ≤28m, decreases by 0.15 for every 10m exceeding" in rule 1); C env This is the environmental correction factor (output by LLM based on real-time weather, time of day, etc., such as C for rainy days). env =1.25).

[0094] Next, normalization and heatmap generation are performed. The initial probability correction values ​​are normalized and mapped to the image space to obtain the second probability heatmap.

[0095] For example, by normalizing the initial probability correction values ​​of all grid cells to the range [0,1], we obtain P. B (i,j) is used to generate a probability heatmap with the same size as the original scene image.

[0096] Figure 4 A second probability heatmap (presented in grayscale) generated according to an embodiment of the present invention is illustrated schematically. This heatmap reflects environmental interactions, specifically the distribution of the intensity of influence of geographical environmental elements (such as oil storage tanks, loading and unloading equipment, dedicated parking spaces, etc.) on target parking behavior. Higher brightness indicates a more significant influence of environmental factors on the area, meaning a stronger driving or attracting effect of environmental factors on target appearance, and vice versa. Taking oil tankers as an example, higher probability values ​​are assigned to the oil tank itself and its adjacent areas (e.g., within the tank area dike, loading and unloading operation area, etc.), reflecting a strong correlation between the probability of target appearance and the spatial location of critical infrastructure—that is, targets tend to move or park within specific functional areas surrounding the oil tank. Figure 4 As shown, the original remote sensing imagery reveals multiple large vertical storage tanks and their ancillary facilities; the second probability heatmap uses grayscale encoding, with higher brightness indicating more significant environmental interactions at that location. Please continue reading. Figure 2 Using a first probability heatmap and a second probability heatmap as weighting factors, a trained object detection model is employed to detect the region of interest. Specifically, this may include: using the object detection model to detect the current image of the region of interest, obtaining the target bounding box and initial confidence score; determining the center point of the bounding box, and obtaining the first probability value and second probability value corresponding to the center point from the first probability heatmap and the second probability heatmap, respectively; weighting and correcting the initial confidence score based on the first probability value and the second probability value to obtain the final confidence score; and assigning the final confidence score to the bounding box and outputting it as the object detection result.

[0097] For example, the probability heatmap P generated based on the target parking pattern A (i,j) and the probability heatmap P generated based on environmental interactions B (i,j) can be applied to the real-time detection process. After using the YOLOv11 network to detect and output the bounding box and its confidence score as preliminary results, the prior probability P of the heatmap location of the bounding box center point can be calculated. A (i,j), P B (i,j) are used as weighting factors. The final confidence level S final It can be adjusted to:

[0098]

[0099] Where S is the initial confidence level, α A α B Adjustable weights (0 < α) A <1, 0<α B <1). In this way, a detection box with low initial confidence but located in a high-probability region will be retained, while a detection box with acceptable initial confidence but located in a low-probability region is more likely to be judged as a false detection and filtered out, thereby improving both recall and precision.

[0100] It should be emphasized that the YOLOv11 used in the detection network can also be replaced with other real-time target detection methods that can output confidence data.

[0101] Please continue reading. Figure 2 The target detection method in embodiments of the present invention may further include a prior knowledge update step based on online data, specifically including: adding the confirmed target detection results to historical observation image data, and updating the first probability heatmap and the second probability heatmap. The confirmed target detection results refer to correct detection results that have been manually reviewed or judged with high confidence. The system can periodically add these correct detection results to the historical observation image database and regenerate the first probability heatmap and the second probability heatmap based on the updated data to replace the original heatmaps. Thus, prior knowledge can adapt to long-term changes in the scene and continuously optimize detection performance without manual intervention.

[0102] Based on the above embodiments, the core advantages and technical effects of the target detection method are reflected in:

[0103] (1) Achieving specific target detection under low imaging quality. This invention reduces the dependence on image information by introducing prior rules of specific targets. Under imaging conditions with low resolution and strong image noise, it can still use the image combined with prior knowledge to perform effective detection.

[0104] (2) Effectively handles detection tasks in complex operating scenarios. When lighting, weather and object occlusion conditions change, the prior knowledge processing method proposed in this invention can guide the algorithm to focus on the detection of specific targets in high-probability areas, thereby significantly improving the detection rate.

[0105] (3) Achieve continuous monitoring of stationary targets. This invention is based entirely on spatial probability and is not affected by the real-time motion state of the target. Therefore, it can effectively detect and monitor targets that have been stationary for a long time, making up for the technical defects of methods based on scene modeling and target tracking.

[0106] (4) Reduce dependence on model complexity and training. Based on the lightweight model, the prior knowledge processing method proposed in this invention can significantly improve the detection performance of the model in specific scenarios and reduce the demand for computing power and training data.

[0107] Based on the above-described target detection method, this invention also provides a target detection device. The following will be combined with... Figure 5 The device is described in detail.

[0108] Figure 5 A schematic block diagram of a target detection device according to an embodiment of the present invention is shown. Figure 5 As shown, the target detection device 500 in this embodiment includes a first generation module 510, a second generation module 520, and a target detection module 530.

[0109] The first generation module 510 is used to generate a first probabilistic heatmap reflecting the parking patterns of targets based on historical observation image data of the region of interest. In one embodiment, the first generation module 510 can be used to perform the operation S110 described above, which will not be repeated here.

[0110] The second generation module 520 is used to generate a second probabilistic heatmap reflecting environmental interactions based on historical observation image data and geographic information system data of the area of ​​interest. In one embodiment, the second generation module 520 can be used to perform the operation S120 described above, which will not be repeated here.

[0111] The target detection module 530 is used to detect the region of interest using a trained target detection model with the first probability heatmap and the second probability heatmap as weighting factors, thereby obtaining the target detection result. In one embodiment, the target detection module 530 can be used to perform the operation S130 described above, which will not be repeated here.

[0112] According to embodiments of the present invention, any plurality of modules among the first generation module 510, the second generation module 520, and the target detection module 530 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules can be combined with at least part of the functionality of other modules and implemented in one module. According to embodiments of the present invention, at least one of the first generation module 510, the second generation module 520, and the target detection module 530 can be at least partially implemented as a hardware circuit, such as a field-programmable gate array, a programmable logic array, a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit, or any other reasonable means of integrating or packaging the circuit, or implemented in any one of the three implementation methods of software, hardware, and firmware, or in a suitable combination of any of these. Alternatively, at least one of the first generation module 510, the second generation module 520, and the target detection module 530 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0113] Figure 6 A block diagram of an electronic device suitable for implementing a target detection method according to an embodiment of the present invention is shown schematically.

[0114] like Figure 6 As shown, an electronic device 600 according to an embodiment of the present invention includes a processor 601, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 602 or a program loaded from storage portion 608 into random access memory (RAM) 603. The processor 601 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit), etc. The processor 601 may also include onboard memory for caching purposes. The processor 601 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0115] RAM 603 stores various programs and data required for the operation of electronic device 600. Processor 601, ROM 602, and RAM 603 are interconnected via bus 604. Processor 601 executes various operations of the method flow according to embodiments of the present invention by executing programs in ROM 602 and / or RAM 603. It should be noted that programs may also be stored in one or more memories other than ROM 602 and RAM 603. Processor 601 may also execute various operations of the method flow according to embodiments of the present invention by executing programs stored in one or more memories.

[0116] According to an embodiment of the present invention, the electronic device 600 may further include an input / output interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the input / output interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube, liquid crystal display, etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0117] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present invention.

[0118] According to embodiments of the present invention, a computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of the present invention, a computer-readable storage medium may include ROM 602 and / or RAM 603 and / or one or more memories other than ROM 602 and RAM 603 described above.

[0119] Embodiments of the present invention also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to enable the computer system to implement the target detection method provided in the embodiments of the present invention.

[0120] When the computer program is executed by the processor 601, it performs the functions defined in the system / apparatus of this invention. According to embodiments of the invention, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0121] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 609, and / or installed from the removable medium 611. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0122] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 609, and / or installed from the removable medium 611. When the computer program is executed by the processor 601, it performs the functions defined in the system of this embodiment of the invention. According to embodiments of the invention, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0123] According to embodiments of the present invention, program code for executing the computer programs provided in the embodiments of the present invention can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0124] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0125] Those skilled in the art will understand that the features described in the various embodiments of the present invention can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention can be combined and / or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

[0126] The embodiments of the present invention have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the invention, and all such substitutions and modifications should fall within the scope of the invention.

Claims

1. A target detection method, characterized in that, include: Based on historical observation image data of the area of ​​interest, a first probability heatmap reflecting the parking pattern of the target is generated. Based on the historical observation image data and the geographic information system data of the area of ​​interest, a second probability heat map reflecting the relationship of environmental effects is generated. Using the first probability heatmap and the second probability heatmap as weighting factors, a trained target detection model is used to detect the region of interest, and the target detection result is obtained.

2. The target detection method according to claim 1, characterized in that, Based on historical observation image data of the area of ​​interest, a first probability heatmap reflecting the parking patterns of targets is generated, including: The region of interest is divided into multiple grid cells, and the initial parking probability of the target in each grid cell is determined based on the historical observation image data. The initial parking probability is corrected based on the spatial distance between each grid cell and the historical target location to obtain the corrected parking probability of each grid cell. The corrected parking probability is normalized and then mapped to the image space to obtain the first probability heatmap.

3. The target detection method according to claim 2, characterized in that, Determining the initial parking probability of the target in each grid cell based on the historical observation image data includes: A weight is assigned to each historical observation frame, and the weight is negatively correlated with the time difference between the historical observation frame and the last frame. The initial parking probability of the grid cell is obtained by weighting and summing the occurrence of the target in each historical observation frame using the weights.

4. The target detection method according to claim 2, characterized in that, The correction of the initial parking probability based on the spatial distance between each grid cell and the historical target location includes: Determine the minimum spatial distance between each grid cell and all historical target locations; An exponential decay coefficient is generated based on the minimum spatial distance, and the exponential decay coefficient decreases as the minimum spatial distance increases; The initial parking probability is corrected using the exponential decay coefficient to obtain the corrected parking probability.

5. The target detection method according to claim 1, characterized in that, Based on the historical observation image data and the geographic information system data of the area of ​​interest, generating a second probabilistic heatmap reflecting environmental interactions includes: The area of ​​interest is divided into multiple grid cells, and points of interest are extracted from the geographic information system data. Using a large language model, semantic association reasoning is performed on the attributes of the points of interest, parking behavior data extracted from the historical observation image data, and dynamic environmental parameters to obtain the association weight, distance decay function, and environmental correction coefficient between each grid cell and each point of interest. The initial probability correction value for each grid cell is determined based on the association weight, the distance decay function, and the environment correction coefficient. The initial probability correction value is normalized and mapped to the image space to obtain the second probability heatmap.

6. The target detection method according to claim 5, characterized in that, Using large language models for semantic association reasoning includes: The attributes of the points of interest, the parking behavior data, and the dynamic environment parameters are used to construct structured prompt words, which are then input into the large language model. The large language model outputs association rules in natural language form, which include the basic weights of each interest point, distance decay conditions, filtering rules for abnormal parking records, and correction rules for dynamic environmental parameters. The association rules are converted into association weights, distance decay functions, and environmental correction coefficients between each grid cell and each point of interest.

7. The target detection method according to claim 5, characterized in that, Determining the initial probability correction value for each of the grid cells includes: For each point of interest, the association weight is multiplied by the distance decay function to obtain the contribution component of the point of interest to the grid cell; The initial probability correction value of the grid cell is obtained by summing the contribution components of all points of interest and multiplying them by the environmental correction coefficient.

8. The target detection method according to claim 1, characterized in that, Using the first probability heatmap and the second probability heatmap as weighting factors, the trained target detection model is used to detect the region of interest, including: The target detection model is used to detect the region of interest to obtain the target detection box and initial confidence score; The center point of the detection frame is determined, and the first probability value and the second probability value corresponding to the center point are obtained from the first probability heatmap and the second probability heatmap, respectively. Based on the first probability value and the second probability value, the initial confidence level is weighted and corrected to obtain the final confidence level; The final confidence level is assigned to the detection box and output as the target detection result.

9. The target detection method according to claim 1, characterized in that, Also includes: The confirmed target detection results are added to the historical observation image data, and the first probability heatmap and the second probability heatmap are updated.

10. A target detection device, characterized in that, include: The first generation module is used to generate a first probability heatmap reflecting the parking pattern of the target based on historical observation image data of the area of ​​interest. The second generation module is used to generate a second probability heatmap reflecting the relationship between environmental effects based on the historical observation image data and the geographic information system data of the area of ​​interest. The target detection module is used to detect the region of interest using the first probability heatmap and the second probability heatmap as weighting factors and a trained target detection model to obtain the target detection result.

Images