Investigation unmanned vehicle target identification method based on computer vision

Abstract

The invention discloses a detection unmanned vehicle target identification method based on computer vision, and relates to the technical field of artificial intelligence. During operation of the system, multi-source heterogeneous vision sensors such as wide-angle visible light, narrow field of view, high resolution and infrared thermal imaging are cooperatively deployed; a lightweight adaptive feature extraction network is introduced, a network structure can flexibly adjust the calculation scale according to the texture complexity of an input image through channel pruning and a dynamic sparse activation mechanism, and the calculation load is effectively reduced while the feature expression capability is ensured; fusing time sequence information through a bidirectional gating circulation unit to generate a target candidate area with motion prior; carrying out weighted fusion on the multi-modal features in the candidate region, and then classifying the multi-modal features; and an online domain adaptive correction mechanism is introduced, and network parameters are finely adjusted in an unsupervised manner according to environmental changes. In addition, occlusion recovery, resource awareness scheduling and offline-online collaborative learning strategies are integrated.

Description

Computer vision-based target recognition method for unmanned reconnaissance vehicles Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a target recognition method for unmanned reconnaissance vehicles based on computer vision. Background Technology

[0002] With the widespread application of intelligent unmanned systems in security, military, and emergency response, computer vision-based reconnaissance drones have become a key platform for performing target recognition tasks in complex environments. These systems rely on onboard cameras and image processing algorithms to perceive the surrounding environment in real time, enabling the automatic detection and classification of personnel, vehicles, or other critical targets. Especially in high-risk scenarios such as urban warfare, disaster search and rescue, or border patrols, reconnaissance drones need to quickly and accurately identify potential threats under conditions of drastic changes in lighting, severe occlusion, or cluttered backgrounds. This places extremely high demands on the robustness, real-time performance, and generalization ability of target recognition algorithms.

[0003] However, existing deep learning-based target recognition methods typically rely on large-scale labeled datasets for training, making it difficult to adapt to unfamiliar environmental distributions in real-world deployments, leading to a significant drop in model performance. Meanwhile, reconnaissance drones, constrained by computational resources and power consumption, struggle to run highly complex visual models, creating an unresolved conflict between recognition accuracy and inference speed. Furthermore, traditional methods often employ independent analysis of static image frames, lacking effective modeling of temporal motion features and failing to fully utilize the target's trajectory and behavioral patterns across consecutive frames. This results in false positives or false negatives under dynamic occlusion or low visibility conditions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a target recognition method for unmanned reconnaissance vehicles based on computer vision, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, this invention employs the following technical solution: a target recognition method for reconnaissance unmanned vehicles based on computer vision, comprising the following specific steps: Step 1: Acquiring multi-source heterogeneous visual data streams: Visible light images, high-resolution local images, and thermal radiation images in the environment are simultaneously acquired using a wide-angle visible light camera, a narrow-field high-resolution camera, and an infrared thermal imaging sensor deployed on the reconnaissance unmanned vehicle. The three types of images are then time-stamp aligned and spatial coordinates unified to form a multi-source heterogeneous visual data stream with spatiotemporal consistency. Step 2: Constructing a lightweight adaptive feature extraction network: A convolutional neural network structure combining channel pruning and dynamic sparse activation is designed. During the inference phase, this network dynamically adjusts the number of activated convolutional channels based on the complexity of the input image. Its backbone network consists of multiple variable-width residual blocks, each containing multiple convolutional layers. The number of output channels of at least one convolutional layer is dynamically adjusted within a preset range to balance computational load and feature representation capability. Step 3: Generating spatiotemporal context-enhanced target candidate regions: Multiple consecutive frames of... Lightweight feature maps are stacked along the time dimension and input into a bidirectional gated recurrent unit module. This module fuses motion cues from historical and future frames through forward and backward state transfer mechanisms, outputting an enhanced spatiotemporal feature map. Based on this feature map, a target candidate region with prior motion trajectory is generated. Step 4 performs multimodal feature fusion and target classification: the feature vectors extracted from visible light, high-resolution, and infrared images are weighted and fused at the corresponding positions of the target candidate region. The weights of each modality feature are allocated according to a preset strategy. The fused feature vector is input into a multilayer fully connected classifier, which outputs the target category and confidence score. Step 5 implements an online domain adaptive correction mechanism: during the operation of the autonomous vehicle, the illumination intensity, contrast, and texture complexity indicators of the current environment image are continuously monitored. When any indicator deviates from the statistical mean of the training set by more than a preset threshold, the lightweight domain adaptive module is triggered. This module fine-tunes the parameters of the last few residual blocks of the feature extraction network through unsupervised adversarial training, enabling the model to quickly adapt to the new environmental distribution. The correction process takes no more than a predetermined time period.

[0006] Preferably, in step 1, the field of view, resolution, and frame rate of the wide-angle visible light camera all meet the preset working requirements; the field of view, resolution, and frame rate of the narrow field of view high-resolution camera also meet another preset working requirement; the thermal sensitivity, spatial resolution, and frame rate of the infrared thermal imaging sensor meet specific performance standards; the three types of sensors achieve high-precision synchronization through hardware trigger signals, and the time alignment error is less than the preset tolerance.

[0007] Preferably, the dynamic sparse activation mechanism of the lightweight adaptive feature extraction network in step 2 is based on the local binary pattern variance calculation of the input image. When the variance is lower than the first preset threshold, only a small number of channels are activated; when the variance is between the first and second preset thresholds, a medium number of channels are activated; and when the variance is higher than the second preset threshold, a large number of channels are activated. This mechanism significantly reduces the computational cost of the network in low-texture scenes while maintaining a high proportion of feature representation capability in high-complexity scenes.

[0008] Preferably, in step 3, the hidden state dimension of the bidirectional gated loop unit module is a preset value, and the fusion of the forward and backward states at time step t adopts the element-wise maximum value operation to retain the most significant motion response; the generated target candidate region not only includes the bounding box coordinates, but also adds velocity vectors and acceleration vectors for subsequent trajectory prediction and occlusion processing.

[0009] Preferably, in step 4, the weight allocation of multimodal feature fusion is dynamically adjusted according to the environmental visibility. When the visibility is lower than a preset distance, the weight of infrared features is increased, the weight of visible light features is reduced accordingly, and the weight of high-resolution features remains unchanged. The visibility is estimated in real time by combining the atmospheric scattering model with the prior of the dark channel of the image, and the update frequency is a preset frequency.

[0010] Preferably, in step 5, the online domain adaptive correction mechanism uses a gradient inversion layer to align the feature distribution. Its adversarial discriminator is composed of a multi-layer fully connected network with a preset number of hidden units. During the fine-tuning process, the first few residual blocks of the backbone network are frozen, and only the convolutional kernel weights and batch normalization parameters of the last few residual blocks are updated. The number of correction iterations is a preset number, and the learning rate is fixed at a preset value.

[0011] Preferably, it also includes a target trajectory smoothing and occlusion recovery module: performing Kalman filtering tracking on the detection results of the same target in consecutive frames; when the target is not detected for several consecutive frames but the predicted position is still within the field of view, occlusion recovery mechanism based on historical trajectory and motion pattern is activated, using the motion vectors of the previous few frames to extrapolate the current position, and prioritizing high-density sampling detection in the area in the next frame.

[0012] Preferably, it also includes a resource-aware scheduling strategy: real-time monitoring of the load rate and memory usage of the autonomous vehicle's central processing unit; when the load rate or memory usage exceeds the corresponding preset threshold, automatically reducing the sampling frequency of the high-resolution camera and lowering the upper limit of the default active channel number of the feature extraction network from the preset maximum value to a lower preset value, in order to maintain the real-time performance of the system and ensure that the overall processing delay is less than the preset time limit.

[0013] Preferably, it also includes an offline-online collaborative learning framework: during task breaks, unlabeled images collected in this task are uploaded to a cloud server, the training set is expanded through semi-supervised learning, and the updated model parameters are downloaded locally periodically to achieve continuous model evolution; local model updates adopt incremental loading, only replacing the changed network layers, avoiding the storage and bandwidth overhead caused by reloading the entire model.

[0014] This invention provides a target recognition method for reconnaissance unmanned vehicles based on computer vision, which has the following beneficial effects: (1) When the system is running, firstly, through the coordinated deployment of multi-source heterogeneous visual sensors such as wide-angle visible light, narrow field of view high resolution and infrared thermal imaging, the environmental information is collected synchronously. And through timestamp alignment and spatial coordinate unification, a visual data foundation with spatiotemporal consistency is constructed, thereby improving the integrity and reliability of the perception data from the source. On this basis, a lightweight adaptive feature extraction network is introduced. Through channel pruning and dynamic sparse activation mechanism, the network structure can flexibly adjust the computational scale according to the texture complexity of the input image, effectively reducing the computational load while ensuring the feature expression capability. Subsequently, the method uses a temporal modeling mechanism to fuse motion information of multiple consecutive frames to generate target candidate regions containing spatial position and motion state, thereby enhancing the perception capability of dynamic targets. Furthermore, through a multimodal feature fusion strategy, the contribution ratio of different modal features is adaptively adjusted according to the environmental visibility to improve the robustness of target recognition under low light or complex weather conditions. Finally, through an online domain adaptive correction mechanism, the model parameters are continuously corrected during actual operation to enable it to quickly adapt to changes in environmental distribution and maintain stable output of recognition performance. Overall, claim 1 constructs a target recognition method for reconnaissance unmanned vehicles that balances real-time performance, accuracy, and environmental adaptability by organically combining multi-source perception, lightweight modeling, spatiotemporal information fusion, and online adaptation.

[0015] (2) Spatiotemporal context modeling: The bidirectional gated loop unit effectively integrates temporal motion information, significantly improving the ability to identify dynamically occluded targets. In test sequences with continuous occlusion exceeding the preset number of frames, the target recapture success rate is much higher than that of the traditional single-frame method. Multimodal complementary enhancement: The weighted fusion of visible light, high resolution and infrared data fully utilizes the advantages of each modality. In rainy and foggy weather, infrared thermal imaging effectively compensates for visible light degradation, significantly improving the recall rate of personnel target recognition, which is significantly higher than that of the single-modal method.

[0016] (3) Online-offline collaborative learning: Through incremental model updates between tasks, the system can continuously optimize performance without interrupting tasks. After multiple task iterations, the accuracy of identifying new camouflaged targets is significantly improved. Occlusion recovery and trajectory smoothing: Kalman filtering and motion extrapolation mechanism effectively reduce target ID switching and trajectory breakage. In dense crowd scenes, the multi-target tracking accuracy (MOTA) reaches a high level, which is significantly better than existing real-time tracking algorithms.

[0017] (4) Overcoming the limitations of static model deployment: Through the integration of online domain adaptive correction mechanism and multimodal perception, the system can maintain stable recognition performance under extreme conditions such as drastic changes in lighting, dense smoke, or low light at night. The average accuracy of target detection reaches a high level in cross-domain testing, which is significantly improved compared with traditional fixed models. Dynamic resource adaptation: The lightweight adaptive feature extraction network and resource perception scheduling strategy work together to dynamically adjust the computing intensity according to the complexity of the scene and the platform load. In typical urban street fighting scenarios, the average inference speed meets the real-time requirements, while the power consumption is greatly reduced. Attached Figure Description

[0018] Figure 1 is a schematic diagram of the steps of the target recognition method for reconnaissance unmanned vehicles based on computer vision according to the present invention; Figure 2 is a schematic diagram of the overall technical solution architecture of the target recognition method for reconnaissance unmanned vehicles based on computer vision proposed in the present invention; Figure 3 is a schematic diagram of the core principle framework of the collaborative work of the lightweight adaptive feature extraction network and the online domain adaptive correction mechanism in the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: This invention provides a target recognition method for reconnaissance unmanned vehicles based on computer vision. Please refer to Figure 1. The method includes the following steps: Step 1: Acquiring multi-source heterogeneous visual data stream: Visible light images, high-resolution local images, and thermal radiation images in the environment are simultaneously acquired by a wide-angle visible light camera, a narrow field-of-view high-resolution camera, and an infrared thermal imaging sensor deployed on the reconnaissance unmanned vehicle. The three types of images are then time-stamp aligned and spatial coordinates unified to form a multi-source heterogeneous visual data stream with spatiotemporal consistency. Step 2: Constructing a lightweight adaptive feature extraction network: A convolutional neural network structure combining channel pruning and dynamic sparse activation is designed. During the inference phase, the number of activated convolutional channels is dynamically adjusted according to the complexity of the input image. The backbone network consists of multiple variable-width residual blocks, each containing multiple convolutional layers. The number of output channels of at least one convolutional layer is dynamically adjusted within a preset range to balance computational load and feature representation capability. Step 3: Generating spatiotemporal context-enhanced target candidate regions: The lightweight features of multiple consecutive frames are used to generate target candidate regions. The images are stacked along the time dimension and input into a bidirectional gated recurrent unit module. This module fuses motion cues from historical and future frames through forward and backward state transfer mechanisms, outputting an enhanced spatiotemporal feature map. Based on this feature map, a target candidate region with prior motion trajectory is generated. Step 4: Perform multimodal feature fusion and target classification: The feature vectors extracted from visible light, high-resolution, and infrared images are weighted and fused at the corresponding positions of the target candidate region. The weights of each modality feature are allocated according to a preset strategy. The fused feature vector is input into a multilayer fully connected classifier, which outputs the target category and confidence score. Step 5: Implement an online domain adaptive correction mechanism: During the operation of the autonomous vehicle, the illumination intensity, contrast, and texture complexity indicators of the current environment image are continuously monitored. When any indicator deviates from the statistical mean of the training set by more than a preset threshold, a lightweight domain adaptive module is triggered. This module fine-tunes the parameters of the last few residual blocks of the feature extraction network through unsupervised adversarial training, enabling the model to quickly adapt to the new environmental distribution. The correction process takes no more than a predetermined time period.

[0021] In this embodiment, by constructing a collaborative acquisition and unified modeling mechanism for multi-source heterogeneous visual data on the reconnaissance unmanned vehicle, consistent expression of different visual modalities in the temporal and spatial dimensions is achieved, significantly improving the integrity and stability of environmental perception in complex scenes. By introducing a lightweight adaptive feature extraction network, the network structure can dynamically adjust the computational scale according to the texture complexity of the input image, effectively reducing computational overhead while ensuring the target feature representation capability, meeting the strict real-time requirements of the unmanned vehicle. By fusing multi-frame temporal information to generate target candidate regions with motion attributes, not only is the ability to identify dynamic targets enhanced, but the accuracy of target localization and motion state estimation is also improved. Through a multi-modal feature adaptive fusion strategy based on environmental visibility, the robustness of target classification under low light, occlusion, or complex weather conditions is improved. And through an online domain adaptive correction mechanism, the model can continuously adapt to changes in environmental distribution during actual operation, reducing the recognition performance degradation caused by scene drift, thereby improving the real-time performance, accuracy, and environmental adaptability of the reconnaissance unmanned vehicle's target recognition.

[0022] Example 2: This example is an explanation of Example 1. Please refer to Figure 1. Specifically: the timestamp alignment is based on the timestamp of the wide-angle visible light image, and the narrow field-of-view high-resolution image and the infrared thermal imaging image are interpolated and aligned, with a time alignment error of less than 5 milliseconds; the spatial coordinates are uniformly projected onto the bird's-eye view plane with the current position of the unmanned vehicle as the origin through a pre-calibrated extrinsic parameter matrix. The visible light and high-resolution images adopt the pinhole camera model, and the infrared image adopts the thermal imaging nonlinear distortion correction model.

[0023] Using the timestamps of images captured by a wide-angle visible light camera as the reference time axis, interpolation alignment is performed on narrow field-of-view high-resolution images and infrared thermal imaging images to ensure that the sampling error of the three types of images on the same time slice does not exceed 5 milliseconds.

[0024] During the process of unifying spatial coordinates, the three types of images are projected onto a bird's-eye view plane with the current position of the unmanned vehicle as the origin through the pre-calibrated extrinsic parameter matrix. The visible light and high-resolution images are mapped using a pinhole camera model, while the infrared thermal imaging images are compensated using a thermal imaging nonlinear distortion correction model.

[0025] In this embodiment, timestamp alignment uses the timestamps of images acquired by a wide-angle visible light camera as the reference time axis. Interpolation alignment is performed on images acquired by a narrow-field-of-view high-resolution camera and an infrared thermal imaging sensor to ensure that the time error between the three types of images at the same moment is less than 5 milliseconds. This method ensures a high degree of synchronization of multimodal images in the time dimension. During spatial coordinate unification, the extrinsic parameter matrices of the three types of sensors relative to the autonomous vehicle's body coordinate system are first obtained through offline calibration. Then, the three types of images are projected onto a bird's-eye view plane with the current position of the autonomous vehicle as the origin. Specifically, the visible light image and the high-resolution image are spatially mapped using a pinhole camera model, while the infrared thermal imaging image is corrected using a thermal imaging nonlinear distortion correction model. Through the above processing, aligned representation of different modal images under a unified spatial coordinate system is achieved.

[0026] By employing the above methods, images of different modalities can be spatially represented in a unified manner, providing a stable data foundation for subsequent feature fusion and target localization.

[0027] Example 3: This example is an explanation of Example 1. Please refer to Figure 2. Specifically: The dynamic sparse activation mechanism of the lightweight adaptive feature extraction network calculates the global texture complexity index based on the local binary pattern variance of the input visible light image. When the index is lower than a first preset threshold, a small number of convolutional channels are activated. When the index is between the first and second preset thresholds, a medium number of convolutional channels are activated. When the index is higher than the second preset threshold, a large number of convolutional channels are activated. Channel activation is achieved through a binary mask generated by a lightweight gating network.

[0028] The system first calculates a global texture complexity index for the input visible light image, which is obtained based on the local binary mode variance or the proportion of high-frequency energy in the image frequency domain. When the texture complexity is lower than a first preset threshold, only a small number of convolutional channels in the residual block are activated; when the texture complexity is between the first and second preset thresholds, a moderate number of convolutional channels are activated; when the texture complexity is higher than the second preset threshold, a larger number of convolutional channels are activated.

[0029] The channel activation process is achieved through a binary mask generated by a lightweight gating network, which significantly reduces the computational load when the scene is simple, while maintaining sufficient feature representation capability in complex scenes, thereby achieving a dynamic balance between performance and efficiency.

[0030] In this embodiment, the feature extraction network adopts a convolutional neural network structure that combines channel pruning and dynamic sparse activation. Its backbone network consists of multiple residual blocks of variable width. Each residual block contains multiple convolutional layers, wherein the number of output channels of at least one convolutional layer can be dynamically adjusted within a preset range according to the texture complexity of the input image.

[0031] Specifically, the system calculates a global texture complexity index based on the local binary pattern variance of the input visible light image. When this index is below a first preset threshold, a lightweight gating network generates a binary mask, activating only a small number of convolutional channels. When the index is between the first and second preset thresholds, a moderate number of convolutional channels are activated. When the index is above the second preset threshold, a larger number of convolutional channels are activated. This dynamic sparse activation mechanism reduces computational cost in low-texture-complexity scenarios while maintaining strong feature extraction capabilities in high-complexity scenarios.

[0032] Example 4: This example is an explanation of Example 1. Please refer to Figures 1 to 3. Specifically: the forward and backward hidden states of the bidirectional gated loop unit module are fused at each time step through element-wise maximum value operation to retain the most significant motion response; the velocity vector of the target candidate region is calculated by displacement difference of the center point of the bounding box between consecutive frames, and the acceleration vector is obtained by further difference of the velocity vector.

[0033] In the bidirectional gated recurrent unit module, the forward hidden state and the backward hidden state are fused at each time step using an element-wise maximum maximization method to retain the most significant motion response information. Based on the enhanced spatiotemporal feature map, the generated target candidate region includes not only spatial bounding box information but also additional velocity and acceleration vectors to describe the target's motion state.

[0034] In addition, the system uses Kalman filtering to smooth the trajectory of the same target in consecutive frames. When the target is not detected for several consecutive frames but the predicted position is still within the field of view, the occlusion recovery mechanism is activated. The system uses historical motion vectors to extrapolate the current position of the target through a uniform acceleration model and performs high-density detection in the prediction area first, thereby improving the success rate of target re-identification under occlusion conditions.

[0035] In this embodiment, feature maps output from a lightweight feature extraction network across multiple consecutive frames are stacked along the time dimension and input into a bidirectional gated recurrent unit (BRU) module. This module fuses the temporal features of the target using a forward and backward state transfer mechanism, while simultaneously utilizing motion information from historical and future frames. At each time step, the forward and backward hidden states of the BRU module are fused through an element-wise maximum value operation to retain the most significant motion response. Based on the fused spatiotemporal feature map, target candidate regions are generated, and bounding box coordinates, a velocity vector, and an acceleration vector are appended to each candidate region. The velocity vector is obtained by subtracting the displacement of the center points of the bounding boxes in consecutive frames, and the acceleration vector is obtained by subtracting the velocity vector again.

[0036] In this way, the target candidate region not only contains spatial information, but also incorporates motion prior information, which helps to improve the stability of dynamic target recognition.

[0037] Example 5: A target recognition method for unmanned reconnaissance vehicles based on computer vision. Referring to Figures 1 to 3, specifically: In step 4, feature vectors for visible light, high resolution, and infrared modes are extracted from the corresponding positions in the target candidate region, and the feature weights for each mode are dynamically allocated based on environmental visibility. Environmental visibility is estimated in real-time using an atmospheric scattering model combined with image dark channel priors. When visibility is below 50 meters, the infrared feature weight is increased and the visible light feature weight is decreased, while the high resolution feature weight remains unchanged; when visibility is not less than 50 meters, the feature weights for each mode are restored to their default values. The weighted and fused feature vectors are input to a multi-layer fully connected classifier, which outputs the target category and confidence level.

[0038] In step 5, the system continuously monitors the illumination intensity, contrast, and texture complexity of the current environmental image. When any of these metrics deviates from the statistical mean of the training set by more than a preset threshold, an online domain adaptive correction mechanism is triggered. This mechanism fine-tunes the parameters of the last few residual blocks of the feature extraction network through unsupervised adversarial training, enabling the model to quickly adapt to new environmental distributions.

[0039] In addition, this embodiment also includes a resource-aware scheduling and offline-online collaborative learning mechanism to ensure that the system can maintain stable and real-time target recognition performance even under conditions of limited computing resources.

[0040] When the autonomous vehicle detects that the light intensity, contrast, or texture complexity of the current environment image deviates from the statistical mean of the training set by more than a preset threshold during operation, the online domain adaptation module is triggered. This module aligns the feature distribution through a gradient inversion layer or a maximum mean difference metric, and only fine-tunes the parameters of a few residual blocks after the feature extraction network, thereby quickly adapting to the new environment without significantly increasing the computational burden.

[0041] Meanwhile, the system monitors the processor load rate and memory usage rate in real time. When any indicator exceeds the preset threshold, it automatically reduces the sampling frequency of the high-resolution camera and limits the upper limit of the number of network active channels to ensure that the overall processing latency meets the real-time requirements.

[0042] During task breaks, the system can also upload the collected unlabeled images to the cloud server for semi-supervised training and update the local model parameters incrementally, enabling offline and online collaborative learning and further improving the model's long-term adaptability.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A target recognition method for unmanned reconnaissance vehicles based on computer vision, characterized in that: Includes the following steps: Step 1: Acquire multi-source heterogeneous visual data stream. Visible light images, high-resolution local images, and thermal radiation images are simultaneously acquired by a wide-angle visible light camera, a narrow field-of-view high-resolution camera, and an infrared thermal imaging sensor deployed on the reconnaissance unmanned vehicle. The three types of images are then time-stamped and their spatial coordinates are unified to form a multi-source heterogeneous visual data stream with spatiotemporal consistency. Step 2: Construct a lightweight adaptive feature extraction network, employing a convolutional neural network structure combining channel pruning and dynamic sparse activation. The backbone network consists of multiple residual blocks of variable width, each containing multiple convolutional layers. The number of output channels of at least one convolutional layer is dynamically adjusted within a preset range based on the texture complexity of the input image. Step 3: Generate spatiotemporal context-enhanced target candidate regions. Lightweight feature maps from multiple consecutive frames are stacked along the time dimension and input into a bidirectional gated recurrent unit module. Motion cues from historical and future frames are fused through forward and backward state transfer mechanisms, outputting an enhanced spatiotemporal feature map. Based on this feature map, a target region containing bounding box coordinates, velocity vectors, and acceleration vectors is generated. Candidate region; Step 4: Perform multimodal feature fusion and target classification. Extract feature vectors of three modalities at the corresponding positions of the target candidate region, and dynamically allocate the feature weights of each modality according to the environmental visibility for weighted fusion. The fused feature vectors are input into a multi-layer fully connected classifier to output the target category and confidence score; Step 5: Implement an online domain adaptive correction mechanism. During the operation of the unmanned vehicle, continuously monitor the illumination intensity, contrast and texture complexity indicators of the current environmental image. When any indicator deviates from the statistical mean of the training set by more than a preset threshold, the lightweight domain adaptive module is triggered. The parameters of the last few residual blocks of the feature extraction network are fine-tuned through unsupervised adversarial training, so that the model can quickly adapt to the new environmental distribution.

2. The target recognition method for unmanned reconnaissance vehicles based on computer vision according to claim 1, characterized in that: The timestamp alignment is based on the timestamp of the wide-angle visible light image, and interpolation alignment is performed on the narrow field-of-view high-resolution image and the infrared thermal imaging image, with a time alignment error of less than 5 milliseconds; the spatial coordinates are uniformly projected onto the bird's-eye view plane with the current position of the unmanned vehicle as the origin through a pre-calibrated extrinsic parameter matrix, wherein the visible light and high-resolution images adopt the pinhole camera model, and the infrared images adopt the thermal imaging nonlinear distortion correction model.

3. The target recognition method for unmanned reconnaissance vehicles based on computer vision according to claim 1, characterized in that: The dynamic sparse activation mechanism of the lightweight adaptive feature extraction network calculates the global texture complexity index based on the local binary pattern variance of the input visible light image. When the index is lower than a first preset threshold, a small number of convolutional channels are activated. When the index is between the first and second preset thresholds, a medium number of convolutional channels are activated. When the index is higher than the second preset threshold, a large number of convolutional channels are activated. Channel activation is achieved through a binary mask generated by a lightweight gating network.

4. The target recognition method for unmanned reconnaissance vehicles based on computer vision according to claim 1, characterized in that: The forward and backward hidden states of the bidirectional gated loop unit module are fused at each time step through an element-wise maximum value operation to retain the most significant motion response; the velocity vector of the target candidate region is calculated by the displacement difference of the center point of the bounding box between consecutive frames, and the acceleration vector is obtained by further differencing the velocity vector.

5. The target recognition method for unmanned reconnaissance vehicles based on computer vision according to claim 1, characterized in that: The weight allocation of the multimodal feature fusion is dynamically adjusted according to the environmental visibility, which is estimated in real time by combining the atmospheric scattering model with the prior of the dark channel of the image. When the visibility is less than 50 meters, the weight of infrared features is increased, the weight of visible light features is decreased, and the weight of high-resolution features remains unchanged. When visibility is not less than 50 meters, the weights of each modal feature are restored to their default values.

6. The target recognition method for unmanned reconnaissance vehicles based on computer vision according to claim 1, characterized in that: The online domain adaptive correction mechanism uses a gradient inversion layer to align feature distributions, and the adversarial discriminator is composed of a multi-layer fully connected network. During the fine-tuning process, the first few residual blocks of the backbone network are frozen, and only the convolutional kernel weights and batch normalization parameters of the last few residual blocks are updated. The number of correction iterations is fixed, the learning rate is fixed, and the correction process takes no more than 200 milliseconds.

7. The target recognition method for unmanned reconnaissance vehicles based on computer vision according to claim 1, characterized in that: It also includes target trajectory smoothing and occlusion recovery steps: Kalman filtering is performed on the detection results of the same target in consecutive frames. The state vector of the Kalman filter includes position, velocity, size and its rate of change. When a target is not detected for three consecutive frames but its predicted position is still within the field of view, the occlusion recovery mechanism is activated. The motion vectors of the previous five frames are used to extrapolate the current position through a uniform acceleration model, and high-density sampling detection is performed in the predicted region in the next frame.

8. The target recognition method for unmanned reconnaissance vehicles based on computer vision according to claim 1, characterized in that: It also includes a resource-aware scheduling step: real-time monitoring of the CPU load and memory usage. When any indicator exceeds the corresponding preset threshold, the sampling frequency of the high-resolution camera is automatically reduced, and the upper limit of the default number of active channels of the feature extraction network is reduced from the preset maximum value to a lower preset value to ensure that the overall processing latency is less than 100 milliseconds.

9. The target recognition method for unmanned reconnaissance vehicles based on computer vision according to claim 1, characterized in that: The dynamic sparse activation mechanism of the lightweight adaptive feature extraction network uses frequency domain analysis to replace local binary mode variance. It determines the texture complexity by calculating the energy ratio of high-frequency components in the input image and dynamically adjusts the number of activation channels accordingly. The weight allocation of the multimodal feature fusion adopts an attention network. The input of this network is the global average pooling features of three modalities, and the output is a normalized weight vector.

10. The target recognition method for unmanned reconnaissance vehicles based on computer vision according to claim 1, characterized in that: It also includes offline and online collaborative learning steps: during task intervals, unlabeled images collected in this task are uploaded to a cloud server, the training set is expanded through semi-supervised learning and an incremental update package is generated, and only the network layer parameters that have changed are replaced locally; the adversarial discriminator in the online domain adaptive correction mechanism is replaced with the maximum mean difference metric, and distribution alignment is achieved by minimizing the maximum mean difference between the source domain and target domain features in the regenerating kernel Hilbert space.

Images