Target detection method and device and electronic equipment

Abstract

The invention relates to a target detection method and device and electronic equipment, and the method comprises the steps: obtaining a target detection image which comprises a target to-be-detected object which is locally shielded; inputting the target detection image into a target feature detection module in the target detection model, and performing horizon feature, two-dimensional feature and three-dimensional feature learning on the target detection image to obtain horizon feature information, two-dimensional feature information and three-dimensional feature information of the target detection image; and inputting the horizon line feature information, the two-dimensional feature information and the three-dimensional feature information into a target object detection module in a target detection model for object detection to obtain a target object detection result of the target detection image. According to the embodiment of the invention, the accuracy of target detection can be improved.

Description

Technical Field

[0001] The present disclosure relates to the field of computer vision technology, and in particular to a target detection method, device, and electronic device. Background Art

[0002] Object detection is a core task in computer vision, aiming to estimate the position and size of objects in an image. In the field of autonomous driving, object detection is the cornerstone of environmental perception. By detecting the position and motion of surrounding vehicles, pedestrians, and obstacles in real time, it enables key functions such as accurately predicting potential collision risks and planning safe driving paths. Therefore, the accuracy and real-time performance of object detection are directly related to the safety and reliability of autonomous driving.

[0003] However, existing object detection technologies still face numerous pressing challenges. Detecting obscured and smaller targets poses challenges in complex road conditions. For example, vehicles obscured by the preceding vehicle and small pedestrians in the distance pose challenges for the algorithm, leading to missed or false detections. In poor weather or lighting conditions, image quality degrades, making feature extraction difficult and significantly reducing detection accuracy. Furthermore, balancing real-time performance with accuracy presents a significant challenge. Due to their high computational complexity, some high-precision algorithms are unable to meet the stringent real-time response requirements of autonomous driving, resulting in insufficient target detection accuracy, inadequate target feature extraction, and low image information utilization. Summary of the Invention

[0004] The present disclosure provides a target detection method, device and electronic device to at least solve the problems of insufficient target detection accuracy, insufficient target feature extraction and low image information utilization in the related art.

[0005] According to a first aspect of an embodiment of the present disclosure, there is provided a target detection method, comprising:

[0006] Acquire a target detection image, wherein the target detection image includes a partially occluded target object to be detected;

[0007] Inputting the target detection image into a target feature detection module in a target detection model, performing horizon feature, two-dimensional feature, and three-dimensional feature learning on the target detection image, and obtaining horizon feature information, two-dimensional feature information, and three-dimensional feature information of the target detection image;

[0008] The horizon feature information, the two-dimensional feature information, and the three-dimensional feature information are input into a target object detection module in the target detection model to perform object detection, and a target object detection result of the target detection image is obtained.

[0009] According to a second aspect of an embodiment of the present disclosure, there is provided an object detection device, including:

[0010] A target detection image acquisition module is used to acquire a target detection image, wherein the target detection image includes a partially obscured target object to be detected;

[0011] a feature information acquisition module, configured to input the target detection image into a target feature detection module in a target detection model, perform horizon feature, two-dimensional feature, and three-dimensional feature learning on the target detection image, and obtain horizon feature information, two-dimensional feature information, and three-dimensional feature information of the target detection image;

[0012] The target object detection result acquisition module is used to input the horizon feature information, the two-dimensional feature information and the three-dimensional feature information into the target object detection module in the target detection model to perform object detection and obtain the target object detection result of the target detection image.

[0013] According to a third aspect of an embodiment of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement a method as described in any one of the first aspects above.

[0014] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute any method described in the first aspect of the embodiment of the present disclosure.

[0015] According to a fifth aspect of the embodiments of the present disclosure, a computer program product containing instructions is provided, which, when executed on a computer, enables the computer to execute any one of the methods described in the first aspect of the embodiments of the present disclosure.

[0016] The technical solutions provided by the embodiments of the present disclosure bring at least the following beneficial effects:

[0017] In the target detection task, a target detection image containing a partially occluded target to be detected is first obtained and then input into the target feature detection module of the target detection model. The target feature detection module extracts the horizon feature information, two-dimensional feature information and three-dimensional feature information of the target detection image by learning the horizon feature, two-dimensional feature information and three-dimensional feature information of the target detection image respectively. These multi-dimensional feature information can comprehensively characterize the content of the target detection image from different angles such as scene reference, plane structure and spatial stereo. Even if the target to be detected is partially occluded, the key information of the target to be detected can be retained through the complementarity of multimodal features. The recognition accuracy in complex occlusion scenes is significantly improved through multi-dimensional feature fusion. Subsequently, the three types of feature information are jointly input into the target object detection module of the target detection model. The target object detection module performs target object detection based on the horizon feature information, two-dimensional feature information and three-dimensional feature information, and can accurately identify and locate the target object in the image. Finally, the target object detection result is output, realizing effective detection of partially occluded targets in complex scenes and improving the accuracy of target detection.

[0018] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute an improper limitation of the present disclosure.

[0020] Figure 1 is a flowchart illustrating a target detection method according to an exemplary embodiment;

[0021] Figure 2 This is a schematic diagram of a process for obtaining horizon feature information, two-dimensional feature information, and three-dimensional feature information of a target detection image according to an exemplary embodiment;

[0022] Figure 3 is a schematic diagram showing a process of obtaining horizon feature information according to an exemplary embodiment;

[0023] Figure 4 is a schematic diagram of a process for obtaining two-dimensional feature information according to an exemplary embodiment;

[0024] Figure 5 is a schematic diagram of a process for obtaining three-dimensional feature information according to an exemplary embodiment;

[0025] Figure 6 is a schematic diagram of a process for obtaining a target detection model according to an exemplary embodiment;

[0026] Figure 7 is a block diagram of a target detection device according to an exemplary embodiment;

[0027] Figure 8 The figure is a block diagram of an electronic device for target detection according to an exemplary embodiment. DETAILED DESCRIPTION

[0028] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0029] It should be noted that the terms "first," "second," and the like in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish between similar and different contents, and are not necessarily used to describe a specific order or precedence. It should be understood that the numbers used in this manner are interchangeable where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of apparatus and methods consistent with certain aspects of the present disclosure as detailed in the appended claims.

[0030] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0031] Figure 1 FIG. 1 is a flow chart of a target detection method according to an exemplary embodiment. Figure 1 As shown, the following steps are included:

[0032] In step S101 , a target detection image is acquired.

[0033] In a specific embodiment, the target detection image includes a partially occluded target object to be detected.

[0034] Exemplarily, the target detection image may be a road environment image acquired during the automatic driving of the vehicle.

[0035] In step S103, the target detection image is input into the target feature detection module in the target detection model, and horizon feature, two-dimensional feature and three-dimensional feature learning are performed on the target detection image to obtain horizon feature information, two-dimensional feature information and three-dimensional feature information of the target detection image.

[0036] In a specific embodiment, before inputting the target detection image into the target feature detection module in the target detection model and performing horizon feature, two-dimensional feature, and three-dimensional feature learning on the target detection image to obtain horizon feature information, two-dimensional feature information, and three-dimensional feature information of the target detection image, the method further includes:

[0037] The target detection image is input into the target feature extraction module, and features of the target detection image are extracted to obtain target image features of the target detection image.

[0038] In a specific embodiment, the target feature extraction module is used to extract multi-level features from the target detection image, from local details to global semantics. The target image features are multi-level semantic representations of the target detection image, integrating information such as the local texture details of the target object, structural relationships between components, and global semantic context.

[0039] In a specific embodiment, the specific module structure of the target feature extraction module can be set in combination with actual application requirements.

[0040] In a specific embodiment, the above-mentioned inputting the target detection image into the target feature extraction module, performing feature extraction on the target detection image, and obtaining the target image features of the target detection image may include: inputting the target detection image into the target feature extraction module, first capturing pixel-level detail features in the shallow network, retaining local information such as target edges and textures; then extracting component-level structural features through mid-layer convolution stacking, establishing spatial associations between the local and overall targets; and finally aggregating global context using large receptive field convolution in the deep network to form a high-level feature expression with semantic consistency. That is, the target feature extraction module performs multi-level feature abstraction on the target detection image through a hierarchical convolutional network, first extracting progressive feature expressions from pixel-level details to semantic-level prototypes layer by layer, then achieving spatial alignment and channel dimension fusion of feature maps at different abstraction levels through a cross-scale feature pyramid architecture, and finally outputting target image features that integrate local structural information and global context associations.

[0041] In a specific embodiment, the specific module structure of the target feature detection module can be set in combination with actual application requirements.

[0042] In a specific embodiment, the target detection image is input into the target feature detection module in the target detection model, and horizon feature, two-dimensional feature, and three-dimensional feature learning are performed on the target detection image. The horizon feature information, two-dimensional feature information, and three-dimensional feature information of the target detection image obtained may include:

[0043] The target image features are input into the target feature detection module, and horizon features, two-dimensional features and three-dimensional features are learned based on the target image features to obtain horizon feature information, two-dimensional feature information and three-dimensional feature information.

[0044] In the above embodiment, during the target detection process, a target detection image containing a partially occluded target to be detected is first obtained, and then input into a target feature extraction module. The target feature extraction module extracts features from the target detection image to obtain multi-level target image features, laying the foundation for subsequent deep feature learning; then, the target image features are input into the target feature detection module to further carry out the learning of horizon features, two-dimensional features and three-dimensional features to form multi-dimensional complementary feature information. Even if the target to be detected is partially occluded, the key semantic information of the target to be detected can be effectively retained.

[0045] In a specific embodiment, the target feature detection module includes a horizon detection submodule, a two-dimensional detection submodule and a three-dimensional detection submodule.

[0046] In a specific embodiment, Figure 2 As shown, the target detection image is input into the target feature detection module in the target detection model, and horizon feature, two-dimensional feature and three-dimensional feature learning are performed on the target detection image. The horizon feature information, two-dimensional feature information and three-dimensional feature information of the target detection image are obtained, including:

[0047] In step S201, the target detection image is input into the horizon detection submodule to perform horizon feature learning to obtain horizon feature information.

[0048] In a specific embodiment, the horizon detection submodule includes a horizon heat map conversion unit and a horizon feature detection unit.

[0049] In a specific embodiment, Figure 3 As shown, the target detection image is input into the horizon detection submodule to perform horizon feature learning, and the horizon feature information obtained includes:

[0050] In step S301, the target detection image is input into the horizon heat map conversion unit to perform heat map conversion to obtain a horizon heat map.

[0051] In one specific embodiment, the horizon heatmap can be a probability density distribution representation of the scene's spatial baseline, presenting the potential geometric orientation of the horizon in the image plane as a pixel-level confidence map. Specifically, the horizon heatmap can be a pixel-level probability density distribution map that quantifies the confidence level of each location in the image plane as the horizon, reflecting the geometric orientation of the scene's three-dimensional spatial baseline.

[0052] In a specific embodiment, the specific unit structure of the horizon heat map conversion unit can be set in combination with actual application requirements.

[0053] In a specific embodiment, the above-mentioned inputting of the target detection image into the horizon heat map conversion unit for heat map conversion to obtain the horizon heat map may include: inputting the target image features into the horizon heat map conversion unit, the horizon heat map conversion unit extracting the feature information required to generate the horizon heat map from the target image features, and then based on the extracted feature information, adaptively weighting the multi-scale features through the cascaded convolution kernel group, combining the spatial attention mechanism to screen local textures and global topological clues that are strongly related to the horizon geometry, and then establishing an implicit mapping relationship between pixel coordinates and horizon parameters, and finally converting the depth continuity information of the three-dimensional scene into a two-dimensional Gaussian probability field, and outputting the horizon heat map.

[0054] In step S303, the horizon heat map is input into the horizon feature detection unit to perform horizon feature learning to obtain horizon feature information.

[0055] In a specific embodiment, the above-mentioned horizon feature information can be structured geometric parameters parsed from the horizon heat map. Specifically, the horizon feature information can be the discrete point feature information of the horizon heat map and the three-dimensional geometric parameterized feature information generated by parsing the discrete key points of the horizon heat map. Exemplarily, the core part of the horizon feature information can be composed of feature information such as directional vector feature information, pitch angle feature information and depth continuity feature information. Among them, the directional vector feature information can describe the spatial extension law of the horizon in the image plane, the pitch angle feature information can quantify the geometric angle between the scene baseline and the camera optical axis, and the depth continuity feature information can model the depth gradient relationship between the near and far scenes through the distribution of key points.

[0056] In a specific embodiment, the specific unit structure of the horizon feature detection unit can be set in combination with actual application requirements.

[0057] In a specific embodiment, the above-mentioned input of the horizon heat map into the horizon feature detection unit to perform horizon feature learning to obtain horizon feature information may include: the horizon feature detection unit locates high-confidence discrete point clusters from the probability distribution of the horizon heat map through sparse key point detection and geometric topological relationship modeling, and fits them into horizon feature information in combination with a continuity optimization algorithm.

[0058] In the above embodiment, the target detection image is converted into a horizon heat map by a horizon heat map conversion unit, which can enhance the semantic significance of the horizon area in the target detection image in the form of probability mapping, so that the horizon stands out from the pixel level in complex backgrounds and occluded scenes, and effectively filter out interference information in non-horizon areas; subsequently, the heat map is subjected to feature learning by a horizon feature detection unit, which can accurately capture the horizontal reference features in the scene and provide key scene space coordinate anchor points for target detection. By combining heat map conversion with feature extraction, the ability to perceive the horizontal structure of the scene is significantly improved, especially in scenes with multiple target occlusions and complex perspectives, which can provide a more reliable global scene prior for the spatial positioning and posture inference of the target to be detected, significantly improving the accuracy of target spatial positioning and the ability to understand the scene context, thereby enhancing the consistency between the position accuracy of the target detection results in three-dimensional space and the semantic understanding of the scene.

[0059] In step S203, the target detection image is input into the two-dimensional detection submodule to perform two-dimensional feature learning to obtain two-dimensional feature information.

[0060] In a specific embodiment, the two-dimensional detection submodule includes a two-dimensional heat map conversion unit and a two-dimensional detection unit.

[0061] In a specific embodiment, Figure 4 As shown, the target detection image is input into the two-dimensional detection submodule to perform two-dimensional feature learning, and the two-dimensional feature information obtained includes:

[0062] In step S401, the target detection image is input into a two-dimensional heat map conversion unit to perform heat map conversion to obtain a two-dimensional heat map.

[0063] In a specific embodiment, the above-mentioned two-dimensional heat map can be a two-dimensional probability density distribution map, which can map the significant area reflecting the geometric shape of the target object to be detected by quantifying the spatial distribution characteristics of the target key points in the image plane at the pixel level confidence level. Among them, the target key points can be the spatial anchor points of the predefined target structural features in the image plane, which are essentially the significant position marks that describe the geometric shape and semantic components of the target object to be detected. Exemplarily, the target key points can be at least one key point such as positioning reference points, component-level key points and semantic association points. Specifically, the positioning reference point can be at least one reference point such as the center point of the target object to be detected (to locate the overall position of the object) and the boundary corner point (to define the detection frame range); the component-level key point can be at least one key point such as the wheel center of the vehicle (to represent the posture), the human body joint (to describe the action) and the building door and window (to analyze the structure); the semantic association point can be at least one association point such as the symmetry axis endpoint of the target object to be detected (to assist deformation analysis) and the occlusion boundary transition point (to support local completion reasoning).

[0064] In a specific embodiment, the specific unit structure of the two-dimensional heat map conversion unit can be set in combination with actual application requirements.

[0065] In a specific embodiment, the above-mentioned inputting of the target detection image into the two-dimensional heat map conversion unit and performing heat map conversion to obtain the two-dimensional heat map may include: inputting the target image features into the two-dimensional heat map conversion unit, the two-dimensional heat map conversion unit extracting the feature information required to generate the two-dimensional heat map from the target image features, and then parsing the extracted feature information through a multi-level feature fusion architecture, using deformable convolution operations to extract spatially sensitive regional response features, combining the channel domain dynamic filtering mechanism to enhance the transmission efficiency of key semantic information, establishing a nonlinear correlation between pixel-level spatial distribution and thermal value in the cross-scale feature interaction process, and finally reconstructing the high-dimensional feature tensor into a two-dimensional probability density field through a deconvolution mapping layer to output a two-dimensional heat map.

[0066] In step S403, the two-dimensional heat map is input into the two-dimensional detection unit to perform two-dimensional feature learning to obtain two-dimensional feature information.

[0067] In a specific embodiment, the above-mentioned two-dimensional feature information can be a geometric parameterized representation generated by decoding the spatial response of the two-dimensional heat map. Exemplarily, the two-dimensional feature information may include feature information such as spatial distribution feature information, semantic attribute feature information and topological structure feature information. Spatial distribution feature information is used to establish a geometric existence representation of the target object to be detected in the target detection image space, and quantify the certainty of positioning. Semantic attribute feature information is used to define the category attribution and morphological essence of the target object to be detected, and achieve discriminability at the semantic level. Topological structure feature information is used to model the dynamic relationship network between the target objects to be detected. Exemplarily, the spatial distribution feature information may include feature information such as the two-dimensional position feature information and boundary diffusion pattern feature information (characterizing the boundary extension trend of the target object to be detected); the semantic attribute feature information may include feature information such as the two-dimensional size feature information of the target object to be detected; and the topological structure feature information may include feature information such as the occlusion coefficient feature information (characterizing the occlusion probability of the target object to be detected).

[0068] In a specific embodiment, the specific unit structure of the two-dimensional detection unit can be set in combination with actual application requirements.

[0069] In a specific embodiment, the above-mentioned input of the two-dimensional heat map into the two-dimensional detection unit, performing two-dimensional feature learning, and obtaining two-dimensional feature information may include: the two-dimensional detection unit extracts the geometric parameters of the high-confidence response area through spatial probability distribution decoding and multi-scale feature aggregation of the two-dimensional heat map, combines non-maximum suppression and continuity optimization, and outputs two-dimensional feature information.

[0070] In the above embodiment, the target detection image is converted into a two-dimensional heat map by a two-dimensional heat map conversion unit, which can enhance the position, contour and structural significance of the target to be detected in the plane space in the form of pixel-level probability mapping, effectively filter out background redundant information and highlight the two-dimensional spatial distribution characteristics of the target to be detected; subsequently, the two-dimensional feature learning of the two-dimensional heat map is performed by the two-dimensional detection unit, which can accurately extract the two-dimensional geometric features of the target to be detected and form two-dimensional feature information containing the target plane structure information. Through the combination of the two-dimensional heat map conversion unit and the two-dimensional detection unit, the perception ability of the two-dimensional spatial morphology of the target to be detected is significantly enhanced, especially in the case of partial occlusion, the key two-dimensional structural features of the visible part of the target to be detected can be retained through the saliency mapping of the two-dimensional heat map, providing clearer plane positioning clues and shape constraints for subsequent target detection, thereby improving the accuracy and anti-occlusion robustness of the two-dimensional position prediction of the target to be detected in complex scenes.

[0071] In step S203, the horizon feature information, the two-dimensional feature information and the target detection image are input into the three-dimensional detection submodule to perform three-dimensional feature learning to obtain three-dimensional feature information.

[0072] In a specific embodiment, the 3D detection submodule includes a length detection unit, a width detection unit, and a depth detection unit, and the depth detection unit includes a first deep learning subunit, a second deep learning subunit, and a target deep learning subunit. The 3D feature information includes target length feature information, target width feature information, and target depth feature information.

[0073] In a specific embodiment, Figure 5 As shown, the horizon feature information, two-dimensional feature information and target detection image are input into the three-dimensional detection submodule to perform three-dimensional feature learning, and the three-dimensional feature information obtained includes:

[0074] In step S501, the target detection image is input into the length detection unit to perform length feature learning to obtain target length feature information.

[0075] In a specific embodiment, the target length characteristic information may be a geometric parameterized representation of the target object to be detected in the X-axis direction in the three-dimensional space, that is, the left and right positions of the target object to be detected in the horizontal direction.

[0076] In a specific embodiment, the specific unit structure of the length detection unit can be set in combination with actual application requirements.

[0077] In a specific embodiment, the above-mentioned inputting the target detection image into the length detection unit to perform length feature learning to obtain the target length feature information may include: inputting the target image feature into the length detection unit, the length detection unit extracting the feature information required to generate the target length feature information from the target image feature, based on the extracted feature information, extracting the local contour response and global scale distribution pattern that are strongly related to the target geometry through a multi-scale convolution kernel group, combining spatial pyramid pooling to fuse structural features under different receptive fields, using an affine transformation layer to model the geometric distortion effects of the target posture and imaging perspective on the length representation, and decoupling the high-dimensional feature map into a scale parameter prediction value through a regression head network, and finally combining the boundary key point offset of the target instance and the prior size constraint matrix to output the target length feature information.

[0078] In step S503, the target detection image is input into the width detection unit to perform width feature learning to obtain target width feature information.

[0079] In a specific embodiment, the target width feature information may be a geometric parameterized representation of the target to be detected along the Y-axis in a three-dimensional space, that is, the height position of the target to be detected in the vertical direction.

[0080] In a specific embodiment, the specific unit structure of the width detection unit can be set in combination with actual application requirements.

[0081] In a specific embodiment, the above-mentioned input of the target detection image into the width detection unit, performing width feature learning, and obtaining target width feature information may include: inputting the target image feature into the width detection unit, the width detection unit extracting the feature information required to generate the target width feature information from the target image feature, based on the extracted feature information, using a direction-selective convolution kernel group to extract the gradient distribution characteristics of the target lateral edge, enhancing the texture continuity perception of the horizontal axis through asymmetric void convolution, using the channel attention mechanism to dynamically screen the multi-level feature responses that are strongly correlated with the width, and at the same time introducing a perspective projection transformation layer to model the lateral scale distortion effect caused by the imaging perspective, separating the target intrinsic width parameters and environmental interference factors through a cascade regression network in the feature decoupling stage, and finally combining the boundary symmetry constraint equation and the lateral key point spatial offset to output the target width feature information.

[0082] In step S505 , the two-dimensional feature information and the preset camera parameters are input into the first depth learning subunit to perform deep learning to obtain first depth feature information.

[0083] In a specific embodiment, the first depth feature information represents projection information of the target object to be detected in the target detection image. The first depth feature information may include horizontal projection feature information of the target object to be detected in the target detection image and vertical projection feature information of the target object to be detected in the target detection image, wherein the horizontal projection feature information represents the lateral distribution characteristics of the target object to be detected in the target detection image plane, and the vertical projection feature information represents the scaling relationship of the target object to be detected in the depth direction.

[0084] In a specific embodiment, the preset camera parameters may include the horizontal focal length of the target camera, the vertical focal length of the target camera, the horizontal coordinates of the principal point of the target camera, and the vertical coordinates of the principal point of the target camera.

[0085] In a specific embodiment, the specific unit structure of the first deep learning sub-unit can be set in combination with actual application requirements.

[0086] In a specific embodiment, the two-dimensional feature information and the preset camera parameters are input into the first deep learning subunit to perform deep learning to obtain the first deep feature information, which may include: the first deep learning subunit calculates the first deep feature information based on the two-dimensional spatial position (u, v) in the two-dimensional feature information and the frame height h in the two-dimensional size, and the formula u b =u, Calculate the coordinates of the target bottom projection point (u b .v b ), and compare it with the preset camera parameters (the horizontal focal length f of the target camera x , the vertical focal length f of the target camera y , the horizontal coordinate c of the target camera principal point u and the vertical coordinate c of the principal point of the target camera v ) into the formula The horizontal projection scale n and vertical projection scale factor m are calculated separately. Then (n, m) is encoded into a geometric feature vector, which is cross-modally fused with the 2D texture features in the 2D feature information through channel splicing. The feature space scale is aligned using affine transformation, and the first deep feature information is finally output that integrates the geometric projection rules and image semantic information.

[0087] In step S507 , the horizon feature information and the preset camera parameters are input into the second depth learning subunit to perform deep learning to obtain second depth feature information.

[0088] In a specific embodiment, the second depth feature information represents the three-dimensional spatial attributes of the target object to be detected, and the second depth feature information may include: fitting the horizon equation y=k using the least squares method based on the discrete coordinate point set in the horizon feature information h x+b h, and get the slope k h and intercept b h ; Then substitute the fixed parameter F=1 into the formula, combined with the horizontal focal length f of the target camera x , the vertical focal length f of the target camera y , calculate the horizontal projection scale parameters Directly assign the balance constraint parameter B = -1, and solve the vertical offset parameter based on the relationship between horizon intercept and focal length. Finally, (A, B, C) is encoded into a geometric feature vector, and the geometric rules and data distribution are fused through the cross-modal feature alignment module to output the second deep feature information.

[0089] In a specific embodiment, the specific unit structure of the second deep learning sub-unit can be set in combination with actual application requirements.

[0090] In step S509 , the first depth feature information, the second depth feature information, and the preset camera parameters are input into a target depth learning subunit to perform depth learning to obtain target depth feature information.

[0091] In a specific embodiment, the target depth feature information is depth feature information obtained by integrating horizon feature information and two-dimensional feature information. The target depth feature information can be a geometric parameterized representation of the target to be detected in the Z-axis direction in three-dimensional space, that is, the distance from the target to be detected to the camera.

[0092] In a specific embodiment, the specific unit structure of the target deep learning sub-unit can be set in combination with actual application requirements.

[0093] In a specific embodiment, the first depth feature information, the second depth feature information and the preset camera parameters are input into the target depth learning subunit to perform deep learning to obtain the target depth feature information, which may include: based on the height H of the target camera, (A, B, C) in the input second depth feature information and (n, m) in the first depth feature information, according to the formula Calculate the global vertical coordinate y glo ; Then combined with the vertical focal length f of the target camera y , the vertical coordinate v of the target bottom projection point b and the vertical coordinate c of the principal point of the target camera v , substitute into the formula Calculate the target depth value; finally output the target depth feature information that integrates geometric constraints and data features.

[0094] In the above embodiment, the length and width feature information of the target are respectively obtained by the length detection unit and the width detection unit, providing basic dimensional data for target analysis. The two-dimensional feature information and preset camera parameters are input into the first deep learning subunit, which generates first depth features representing the target projection information based on the plane features and camera parameters, accurately describing the target's plane projection characteristics. The horizon feature information and preset camera parameters are input into the second deep learning subunit, which mines second depth features representing the target's three-dimensional spatial properties. Finally, the two types of depth feature information and the preset camera parameters are input into the target deep learning subunit, achieving deep fusion of the horizon features and the two-dimensional features to obtain comprehensive target depth feature information. This not only considers the plane projection details of the target to be detected, but also infers the three-dimensional spatial relationship based on the horizon feature information. This achieves deep complementary optimization of feature information, more comprehensively and accurately understands the target to be detected's true state in the three-dimensional scene, improves target detection accuracy and spatial structure perception capabilities in complex scenes, and provides rich and reliable depth feature support for subsequent target positioning, recognition, and scene understanding.

[0095] In a specific embodiment, the specific module structures of the horizon detection submodule, the two-dimensional detection submodule and the three-dimensional detection submodule can be set in combination with actual application requirements.

[0096] In the above embodiment, the target detection image is input into the horizon detection submodule to obtain horizon feature information, which can determine the scene horizontal reference; the two-dimensional detection submodule is input to obtain two-dimensional feature information, which is used to describe the target plane structure; and the three-dimensional detection submodule is then used to fuse the horizon feature information, two-dimensional feature information and target detection image to perform three-dimensional feature learning. This enables the three-dimensional detection submodule to understand the three-dimensional properties of the target to be detected from a three-dimensional spatial level on the basis of grasping the scene horizontal reference and target plane details, and improve the perception ability of the spatial position, posture and other information of the target to be detected, providing richer information support for accurate target detection and scene analysis, and enhancing the system's understanding and processing efficiency of complex scenes.

[0097] In a specific embodiment, the three-dimensional feature information may also include three-dimensional spatial position feature information of the target object to be detected and direction feature information of the target object to be detected. Specifically, the three-dimensional detection submodule may also include a three-dimensional spatial position detection unit and a direction detection unit.

[0098] In a specific embodiment, inputting the target detection image into the feature detection module in the target detection model, performing three-dimensional feature learning on the target detection image, and obtaining three-dimensional feature information may include: inputting the target image features into the three-dimensional space position detection unit, performing three-dimensional space position detection, and obtaining three-dimensional space position feature information; inputting the target image features into the direction detection unit, performing direction detection, and obtaining direction feature information.

[0099] Specifically, inputting the target image features into a three-dimensional space position detection unit to perform three-dimensional space position detection to obtain three-dimensional space position feature information may include: inputting the target image features into the three-dimensional space position detection unit, the three-dimensional space position detection unit extracting feature information required to generate three-dimensional space position feature information from the target image features, then extracting spatial orientation features through deformable convolution, combining camera parameters for coordinate system mapping, using a decoupling head to separate absolute coordinates and height offset, and finally integrating point cloud and texture features to output three-dimensional space position feature information.

[0100] Specifically, inputting the target image features into a direction detection unit to perform direction detection to obtain directional feature information may include: inputting the target image features into the direction detection unit, the direction detection unit extracting feature information required to generate directional feature information from the target image features, and then performing polar coordinate conversion and spatiotemporal attention to separate motion trends and background interference, using a decoupling head to fuse multi-scale contexts and model angle distribution, and combining geometric constraints to output directional feature information.

[0101] In step S105 , the horizon feature information, the two-dimensional feature information, and the three-dimensional feature information are input into a target object detection module in a target detection model to perform object detection, thereby obtaining a target object detection result of a target detection image.

[0102] In practical applications, autonomous driving decisions can be made based on target detection results.

[0103] In a specific embodiment, Figure 6 As shown in Figure 2, the above target detection model is obtained in the following way:

[0104] In step S601, a sample detection image set, an original image corresponding to each sample detection image in the sample detection image set, and a preset object detection result corresponding to each sample detection image are obtained, where each sample detection image includes a partially occluded sample object to be detected.

[0105] In step S603, a current sample detection image is determined from the sample detection image set.

[0106] In step S605, the current sample detection image is input into the feature extraction module to be trained in the detection model to be trained, and features are extracted from the current sample detection image to obtain sample image features of the current sample detection image.

[0107] In step S607, the sample image features are input into the feature detection module to be trained and the object detection module to be trained in the detection model to be trained. On the basis of learning the horizon features, two-dimensional features and three-dimensional features of the sample detection image, object detection is performed to obtain the sample object detection result of the sample detection image.

[0108] In a specific embodiment, the above-mentioned sample image features are input into the feature detection module to be trained and the object detection module to be trained in the detection model to be trained, and object detection is performed on the basis of learning the horizon features, two-dimensional features and three-dimensional features of the sample detection image to obtain the specific refinement of the sample object detection result of the sample detection image. Please refer to the above-mentioned step S103, input the target detection image into the target feature detection module in the target detection model, learn the horizon features, two-dimensional features and three-dimensional features of the target detection image, and obtain the horizon feature information, two-dimensional feature information and three-dimensional feature information of the target detection image; and step S105, input the horizon feature information, two-dimensional feature information and three-dimensional feature information into the target object detection module in the target detection model to perform object detection, and obtain the target object detection result of the target detection image, which will not be repeated here.

[0109] In step S609 , the sample image features are input into a preset image restoration module, and image restoration is performed based on the sample image features to obtain a sample restored image of the sample detection image.

[0110] In step S611, based on the sample object detection result, the preset object detection result corresponding to the current sample detection image, the sample restored image and the original image corresponding to the current sample detection image, the detection model to be trained is trained to obtain a target detection model.

[0111] In a specific embodiment, the above-mentioned training of the detection model to be trained based on the sample object detection result, the preset object detection result corresponding to the current sample detection image, the sample restoration image and the original image corresponding to the current sample detection image to obtain the target detection model may include: determining the feature extraction loss information of the feature extraction module to be trained based on the sample restoration image and the original image corresponding to the current sample detection image, and updating the model parameters of the feature extraction module to be trained based on the feature extraction loss information; determining the feature detection loss information corresponding to the feature detection module to be trained and the result detection loss information corresponding to the object detection module to be trained based on the sample object detection result and the preset object detection result corresponding to the current sample detection image; and updating the model parameters of the feature detection module to be trained and the object detection module to be trained based on the feature detection loss information and the result detection loss information. And based on the updated feature extraction module to be trained, the feature detection module to be trained and the object detection module to be trained, repeat the above-mentioned iterative training steps of determining the current sample detection image from the sample detection image set to updating the model parameters of the feature extraction module to be trained and the model parameters of the feature detection module to be trained and the object detection module to be trained based on the feature detection loss information and the result detection loss information until the preset convergence conditions are met, and the detection model to be trained that meets the preset convergence conditions is used as the target detection model.

[0112] In a specific embodiment, the current sample detection image may be a detection image in this cyclic iterative training process; specifically, the current sample detection image may be part of the detection images in the sample detection image set; specifically, the current sample detection image may be randomly selected, or the current sample detection image may be determined in combination with a preset selection rule.

[0113] In a specific embodiment, the sample object detection result corresponding to the current sample detection image, the preset object detection result corresponding to the current sample detection image, the sample restoration image corresponding to the current sample detection image, and the original image corresponding to the current sample detection image can be substituted into a preset loss function to determine the feature extraction loss information corresponding to the feature extraction module to be trained, the feature detection loss information corresponding to the feature detection module to be trained, and the result detection loss information corresponding to the object detection module to be trained. The feature extraction loss information can characterize the feature extraction performance of the current feature extraction module to be trained, the feature detection loss information can characterize the feature detection performance of the current feature detection module to be trained, and the result detection loss information can characterize the object detection performance of the current object detection module to be trained. Specifically, the preset loss function can be set in combination with actual applications.

[0114] In the above embodiment, during the target detection model training process, a sample detection image set containing a partially occluded sample object to be detected is obtained, and the original image and the preset object detection results are combined to sequentially perform feature extraction, multi-dimensional feature learning (horizontal features, two-dimensional features, three-dimensional features), and object detection on the current sample detection image. Simultaneously, a preset image restoration module is used to restore the sample detection image based on the sample image features to form a sample restoration image. This training method effectively guides the trained detection model to learn the multi-dimensional feature representation of the target in the occluded scene by comparing the sample object detection results with the preset object detection results, and combining the difference feedback between the sample restoration image and the original image. The image restoration process strengthens the trained feature extraction module's ability to reason about the structure and complete the details of the occluded area. Ultimately, the trained target detection model can accurately identify and locate partially occluded targets in complex occluded scenes through multimodal feature fusion and image structure constraints, significantly improving the detection robustness, semantic understanding accuracy, and spatial positioning accuracy of the occluded target, effectively solving the problem of insufficient feature utilization and insufficient detection accuracy of traditional detection models in occluded scenes.

[0115] In a specific embodiment, the sample detection image set is obtained in the following manner:

[0116] The original image corresponding to each sample detection image is randomly blocked to obtain a sample detection image set.

[0117] In a specific embodiment, the random occlusion may be a occlusion method that covers different occlusion positions, occlusion areas, occlusion shapes, and the like.

[0118] In the above embodiment, by randomly occluding the original image corresponding to each sample detection image to generate a sample detection image set, it is possible to effectively simulate the complex situation of partial occlusion of the target in the real scene, provide a variety of occlusion samples for the target detection model training, and enable the detection model to be trained to fully contact the target features in various occlusion scenarios during the training stage, strengthen the ability to extract the residual visible features of the occluded sample to be detected objects and the contextual semantic reasoning ability, which can significantly improve the richness of the training data, and force the detection model to be trained to learn to use the texture, contour, spatial position relationship and other information of the unoccluded part of the sample to be detected object to infer the structural features of the occluded area, avoiding the detection model to be trained from relying on complete target features for detection, thereby enhancing the model's detection robustness for partially occluded target to be detected objects in actual applications, enabling it to more accurately identify and locate target to be detected objects in complex occlusion environments, and effectively solving the problem of insufficient generalization ability of traditional detection models due to the lack of real occlusion scenes in training data.

[0119] Figure 7FIG. 1 is a block diagram of a target detection device according to an exemplary embodiment. Figure 7 , the device comprises:

[0120] The target detection image acquisition module 710 is used to acquire a target detection image, wherein the target detection image includes a partially occluded target object to be detected;

[0121] A feature information acquisition module 720 is configured to input the target detection image into the target feature detection module in the target detection model, perform horizon feature, two-dimensional feature, and three-dimensional feature learning on the target detection image, and obtain horizon feature information, two-dimensional feature information, and three-dimensional feature information of the target detection image;

[0122] The target object detection result acquisition module 730 is used to input the horizon feature information, the two-dimensional feature information and the three-dimensional feature information into the target object detection module in the target detection model to perform object detection and obtain the target object detection result of the target detection image.

[0123] In an optional embodiment, the feature detection module includes a horizon detection submodule, a two-dimensional detection submodule, and a three-dimensional detection submodule. The feature information acquisition module 720 includes:

[0124] The horizon feature information acquisition unit is used to input the target detection image into the horizon detection submodule, perform horizon feature learning, and obtain horizon feature information.

[0125] The two-dimensional feature information acquisition unit is used to input the target detection image into the two-dimensional detection submodule, perform two-dimensional feature learning, and obtain two-dimensional feature information.

[0126] The three-dimensional feature information acquisition unit is used to input the horizon feature information, the two-dimensional feature information and the target detection image into the three-dimensional detection submodule to perform three-dimensional feature learning to obtain three-dimensional feature information.

[0127] In an optional embodiment, the three-dimensional detection submodule includes a length detection unit, a width detection unit, and a depth detection unit, and the depth detection unit includes a first deep learning subunit, a second deep learning subunit, and a target deep learning subunit. The three-dimensional feature information includes target length feature information, target width feature information, and target depth feature information. The three-dimensional feature information acquisition unit includes:

[0128] The target length feature information acquisition subunit is used to input the target detection image into the length detection unit, perform length feature learning, and obtain target length feature information.

[0129] The target width feature information acquisition subunit is used to input the target detection image into the width detection unit, perform width feature learning, and obtain target width feature information.

[0130] The first depth feature information acquisition subunit is used to input the two-dimensional feature information and preset camera parameters into the first depth learning subunit to perform deep learning to obtain first depth feature information.

[0131] The second depth feature information acquisition subunit is used to input the horizon feature information and preset camera parameters into the second depth learning subunit to perform deep learning to obtain second depth feature information.

[0132] The target depth feature information acquisition subunit is used to input the first depth feature information, the second depth feature information and the preset camera parameters into the target depth learning subunit to perform deep learning to obtain the target depth feature information.

[0133] In an optional embodiment, the horizon detection submodule includes a horizon heat map conversion unit and a horizon feature detection unit. The horizon feature information acquisition unit includes:

[0134] The horizon heat map acquisition subunit is used to input the target detection image into the horizon heat map conversion unit, perform heat map conversion, and obtain the horizon heat map.

[0135] The horizon feature information acquisition subunit is used to input the horizon heat map into the horizon feature detection unit, perform horizon feature learning, and obtain horizon feature information.

[0136] In an optional embodiment, the two-dimensional detection submodule includes a two-dimensional heat map conversion unit and a two-dimensional detection unit. The two-dimensional feature information acquisition unit includes:

[0137] The two-dimensional heat map acquisition subunit is used to input the target detection image into the two-dimensional heat map conversion unit, perform heat map conversion, and obtain a two-dimensional heat map.

[0138] The two-dimensional feature information acquisition subunit is used to input the two-dimensional heat map into the two-dimensional detection unit, perform two-dimensional feature learning, and obtain two-dimensional feature information.

[0139] In an optional embodiment, the target detection model further includes a target feature extraction module. Before the feature information acquisition module 720, the device further includes:

[0140] The target image feature acquisition module is used to input the target detection image into the target feature extraction module, perform feature extraction on the target detection image, and obtain the target image features of the target detection image.

[0141] The feature information acquisition module 720 includes:

[0142] The feature information acquisition unit is used to input the target image features into the feature detection module, perform horizon feature, two-dimensional feature and three-dimensional feature learning based on the target image features, and obtain horizon feature information, two-dimensional feature information and three-dimensional feature information.

[0143] In an optional embodiment, the target detection model is obtained using the following modules:

[0144] The information acquisition module is used to obtain a sample detection image set, the original image corresponding to each sample detection image in the sample detection image set, and the preset object detection result corresponding to each sample detection image. Each sample detection image includes a partially obscured sample object to be detected.

[0145] The current sample detection image determination module is used to determine the current sample detection image from the sample detection image set.

[0146] The sample image feature acquisition module is used to input the current sample detection image into the to-be-trained feature detection module in the to-be-trained detection model, perform feature extraction on the current sample detection image, and obtain the sample image features of the current sample detection image.

[0147] The sample object detection result acquisition module is used to input the sample image features into the to-be-trained feature detection module and the to-be-trained object detection module in the to-be-trained detection model, perform object detection on the basis of learning the horizon features, two-dimensional features and three-dimensional features of the sample detection image, and obtain the sample object detection results of the sample detection image.

[0148] The sample restoration image acquisition module is used to input the sample image features into the preset image restoration module, perform image restoration based on the sample image features, and obtain a sample restoration image of the sample detection image.

[0149] The target detection model determination module is used to train the detection model to be trained based on the sample object detection results, the preset object detection results corresponding to the current sample detection image, the sample restored image and the original image corresponding to the current sample detection image to obtain the target detection model.

[0150] In an optional embodiment, the sample detection image set is obtained using the following modules:

[0151] The sample detection image set acquisition module is used to randomly block the original image corresponding to each sample detection image to obtain the sample detection image set.

[0152] Figure 8 is a block diagram of an electronic device for target detection according to an exemplary embodiment. The electronic device may be a server, and its internal structure diagram may be as shown in FIG. Figure 8As shown. The electronic device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the electronic device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a target detection method is implemented. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the electronic device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the electronic device, or an external keyboard, touchpad or mouse, etc.

[0153] Those skilled in the art will understand that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the scheme of the present disclosure, and does not constitute a limitation on the electronic device to which the scheme of the present disclosure is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0154] In an exemplary embodiment, an electronic device is further provided, including: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the target detection method as in the embodiment of the present disclosure.

[0155] In an exemplary embodiment, a computer-readable storage medium is further provided. When instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the target detection method in the embodiment of the present disclosure.

[0156] In an exemplary embodiment, a computer program product including instructions is further provided. When the computer program product is run on a computer, the computer is caused to perform the target detection method in the embodiment of the present disclosure.

[0157] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, which can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0158] Other embodiments of the present disclosure will readily occur to those skilled in the art after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the following claims.

[0159] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A target detection method, characterized in that: include: Acquire a target detection image, wherein the target detection image includes a partially occluded target object to be detected; Inputting the target detection image into a target feature detection module in a target detection model, performing horizon feature, two-dimensional feature, and three-dimensional feature learning on the target detection image, and obtaining horizon feature information, two-dimensional feature information, and three-dimensional feature information of the target detection image; The horizon feature information, the two-dimensional feature information, and the three-dimensional feature information are input into a target object detection module in the target detection model to perform object detection, and a target object detection result of the target detection image is obtained.

2. The method according to claim 1, characterized in that The feature detection module includes a horizon detection submodule, a two-dimensional detection submodule, and a three-dimensional detection submodule. The target detection image is input into the target feature detection module in the target detection model, and horizon feature, two-dimensional feature, and three-dimensional feature learning are performed on the target detection image to obtain horizon feature information, two-dimensional feature information, and three-dimensional feature information of the target detection image. Inputting the target detection image into the horizon detection submodule to perform horizon feature learning to obtain the horizon feature information; Inputting the target detection image into the two-dimensional detection submodule to perform two-dimensional feature learning to obtain the two-dimensional feature information; The horizon feature information, the two-dimensional feature information and the target detection image are input into the three-dimensional detection submodule to perform three-dimensional feature learning to obtain the three-dimensional feature information.

3. The method according to claim 2, characterized in that The three-dimensional detection submodule includes a length detection unit, a width detection unit, and a depth detection unit. The depth detection unit includes a first depth learning subunit, a second depth learning subunit, and a target depth learning subunit. The three-dimensional feature information includes target length feature information, target width feature information, and target depth feature information. The horizon feature information, the two-dimensional feature information, and the target detection image are input into the three-dimensional detection submodule to perform three-dimensional feature learning. The three-dimensional feature information obtained includes: Inputting the target detection image into the length detection unit to perform length feature learning to obtain the target length feature information; Inputting the target detection image into the width detection unit to perform width feature learning to obtain target width feature information; Inputting the two-dimensional feature information and preset camera parameters into the first deep learning subunit to perform deep learning to obtain first depth feature information, where the first depth feature information represents projection information of the target object to be detected in the target detection image; Inputting the horizon feature information and the preset camera parameters into the second deep learning subunit to perform deep learning to obtain second depth feature information, where the second depth feature information represents the three-dimensional spatial attributes of the target object to be detected; The first depth feature information, the second depth feature information and the preset camera parameters are input into the target depth learning subunit to perform deep learning to obtain the target depth feature information, where the target depth feature information is depth feature information obtained by integrating the horizon feature information and the two-dimensional feature information.

4. The method according to claim 2, characterized in that The horizon detection submodule includes a horizon heat map conversion unit and a horizon feature detection unit. The target detection image is input into the horizon detection submodule to perform horizon feature learning to obtain the horizon feature information. Inputting the target detection image into the horizon heat map conversion unit to perform heat map conversion to obtain a horizon heat map; The horizon heat map is input into the horizon feature detection unit to perform horizon feature learning to obtain the horizon feature information.

5. The method according to claim 2, characterized in that: The two-dimensional detection submodule includes a two-dimensional heat map conversion unit and a two-dimensional detection unit. The target detection image is input into the two-dimensional detection submodule to perform two-dimensional feature learning to obtain the two-dimensional feature information. Inputting the target detection image into the two-dimensional heat map conversion unit to perform heat map conversion to obtain a two-dimensional heat map; The two-dimensional heat map is input into the two-dimensional detection unit to perform two-dimensional feature learning to obtain the two-dimensional feature information.

6. The method according to claim 1, characterized in that The target detection model further includes a target feature extraction module. Before inputting the target detection image into the target detection model and performing horizon feature, two-dimensional feature, and three-dimensional feature learning on the target detection image to obtain horizon feature information, two-dimensional feature information, and three-dimensional feature information of the target detection image, the method further includes: Inputting the target detection image into the target feature extraction module, performing feature extraction on the target detection image, and obtaining target image features of the target detection image; Inputting the target detection image into a target feature detection module in a target detection model, performing horizon feature, two-dimensional feature, and three-dimensional feature learning on the target detection image, and obtaining horizon feature information, two-dimensional feature information, and three-dimensional feature information of the target detection image includes: The target image features are input into the feature detection module, and horizon features, two-dimensional features and three-dimensional features are learned based on the target image features to obtain the horizon feature information, the two-dimensional feature information and the three-dimensional feature information.

7. The method according to claim 1, characterized in that: The target detection model is obtained in the following way: Obtaining a sample detection image set, an original image corresponding to each sample detection image in the sample detection image set, and a preset object detection result corresponding to each sample detection image, wherein each sample detection image includes a partially occluded sample object to be detected; Determining a current sample detection image from the sample detection image set; Inputting the current sample detection image into a to-be-trained feature extraction module in the to-be-trained detection model, performing feature extraction on the current sample detection image, and obtaining sample image features of the current sample detection image; Inputting the sample image features into the to-be-trained feature detection module and the to-be-trained object detection module in the to-be-trained detection model, performing object detection on the basis of learning horizon features, two-dimensional features, and three-dimensional features of the sample detection image, and obtaining a sample object detection result of the sample detection image; Inputting the sample image features into a preset image restoration module, performing image restoration based on the sample image features, and obtaining a sample restored image of the sample detection image; Based on the sample object detection result, the preset object detection result corresponding to the current sample detection image, the sample restored image and the original image corresponding to the current sample detection image, the detection model to be trained is trained to obtain the target detection model.

8. The method according to claim 7, characterized in that: The sample detection image set is obtained in the following manner: The original image corresponding to each sample detection image is randomly blocked to obtain the sample detection image set.

9. A target detection device, characterized in that: include: A target detection image acquisition module is used to acquire a target detection image, wherein the target detection image includes a partially obscured target object to be detected; a feature information acquisition module, configured to input the target detection image into a target feature detection module in a target detection model, perform horizon feature, two-dimensional feature, and three-dimensional feature learning on the target detection image, and obtain horizon feature information, two-dimensional feature information, and three-dimensional feature information of the target detection image; The target object detection result acquisition module is used to input the horizon feature information, the two-dimensional feature information and the three-dimensional feature information into the target object detection module in the target detection model to perform object detection and obtain the target object detection result of the target detection image.

10. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the target detection method according to any one of claims 1 to 8.

Images