Automatic driving perception enhancement system for police unmanned inspection vehicle

Abstract

The invention discloses an automatic driving perception enhancement system for a police unmanned inspection vehicle, which belongs to the technical field of automatic driving and comprises a sensing data acquisition module, a self-adaptive perception enhancement module, a target identification module and a risk early warning module. According to the invention, adaptive perception enhancement is carried out by adopting an environment perception adaptive adjustment method, different inspection environments are perceived in real time, and the image and point cloud quality is adjusted as required, so that a stable perception effect is kept under a complex meteorological condition, and reliable data support is provided for subsequent target identification; target recognition is carried out by adopting an environment perception driven multi-stage target recognition method, different types of targets are comprehensively recognized, key targets and abnormal conditions can still be stably recognized in a complex environment, and the recognition reliability of the police unmanned inspection vehicle in all weather and multiple scenes is remarkably improved; risk early warning is carried out by comprehensively analyzing environmental risks and target anomalies, abnormal behaviors in a complex environment are identified, and the risks are dynamically evaluated in combination with the current environmental state.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically referring to an autonomous driving perception enhancement system for police unmanned patrol vehicles. Background Technology

[0002] The autonomous driving perception enhancement system for police unmanned patrol vehicles uses artificial intelligence technology to analyze multi-source data during the patrol process. It aims to achieve intelligent perception and adjustment of abnormal environments such as low light, rain, snow, and fog, and enhance the ability of unmanned patrol vehicles to identify suspicious targets during actual patrols, thereby improving the task execution efficiency of the autonomous driving system in complex police scenarios.

[0003] However, in the current process of enhancing perception for autonomous driving, there are technical problems such as poor adaptability to complex environments, large fluctuations in the quality of perceived images, which leads to a significant decrease in the accuracy of target recognition under adverse weather conditions and easy omission of key targets; there are technical problems such as the accuracy of target recognition being greatly affected by environmental interference, and the difficulty in recognizing small or fast-moving targets, which affects the inspection effect; there are technical problems such as untimely response to sudden environmental changes and suspicious target behavior, and the failure to comprehensively consider environmental background and target behavior in the risk assessment process, resulting in incomplete early warning information. Summary of the Invention

[0004] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides an autonomous driving perception enhancement system for police unmanned patrol vehicles. Addressing the technical problems of poor adaptability to complex environments and large fluctuations in perceived image quality in existing autonomous driving perception enhancement processes, leading to a significant decrease in target recognition accuracy and the easy omission of key targets in adverse weather conditions, this solution creatively employs an adaptive environmental perception adjustment method for adaptive perception enhancement. This method perceives different patrol environments in real time and adjusts image and point cloud quality as needed, thereby maintaining stable and clear perception effects under complex weather conditions such as nighttime, rain, snow, and fog, providing reliable data support for subsequent target recognition and effectively meeting the intelligent patrol requirements of police scenarios for all-weather, high-accuracy, and low-false-judgment capabilities. Furthermore, addressing the technical problems of target recognition accuracy being greatly affected by environmental interference and difficulty in identifying small or fast-moving targets in existing autonomous driving perception enhancement processes, which impacts patrol effectiveness, this solution creatively… This solution innovatively employs a multi-level target recognition method driven by environmental perception to comprehensively identify targets of different types, sizes, and behavioral states. It can intelligently adjust its recognition strategy based on environmental conditions, ensuring stable identification of key targets and anomalies even in complex environments. This significantly improves the reliability of police unmanned patrol vehicles in all-weather, multi-scenario environments. Addressing the technical issues in existing autonomous driving perception enhancement processes, such as untimely responses to sudden environmental changes and suspicious target behavior, and incomplete early warning information due to the failure to comprehensively consider environmental background and target behavior during risk assessment, this solution creatively provides risk warnings through comprehensive analysis of environmental risks and target anomalies. This achieves real-time perception and intelligent response to potentially high-risk areas, not only detecting abnormal behavior in complex environments in advance but also dynamically assessing risk levels based on the current environmental state. This effectively enhances the risk identification capabilities and handling efficiency of police unmanned patrol vehicles in various environments.

[0005] The technical solution adopted by the present invention is as follows: The present invention provides an autonomous driving perception enhancement system for police unmanned patrol vehicles, comprising: a sensor data acquisition module, an adaptive perception enhancement module, a target recognition module, and a risk warning module;

[0006] The sensor data acquisition module is used for sensor data acquisition. Through sensor data acquisition, multi-source sensor data is obtained, and the multi-source sensor data is sent to the adaptive perception enhancement module.

[0007] The adaptive perception enhancement module is used for adaptive perception enhancement. Based on multi-source sensor data, it adopts an environmental perception adaptive adjustment method to perform adaptive perception enhancement, obtains image point cloud enhancement features, and sends the image point cloud enhancement features to the target recognition module.

[0008] The target recognition module is used for target recognition. Based on the enhanced features of the image point cloud, it uses an environment-aware-driven multi-level target recognition method to perform target recognition, obtain target recognition reference data, and send the target recognition reference data to the risk warning module.

[0009] The risk warning module is used for risk warning. Based on target identification reference data, it conducts risk warning by comprehensively analyzing environmental risks and target anomalies, and generates warning information.

[0010] Furthermore, the sensor data acquisition specifically involves, during the patrol process of the police unmanned patrol vehicle, the on-board multimodal sensor system to collect image frame data, point cloud frame data, inertial measurement data, and meteorological data in real time, and to perform time alignment through a timestamp synchronization mechanism to obtain synchronized multi-source data; then, based on the spatial calibration relationship between the sensors, the synchronized multi-source data is spatially aligned to obtain multi-source sensor data.

[0011] Furthermore, the adaptive perception enhancement is used to perceive the environmental state during the inspection process. The adaptive perception enhancement is performed using an environmental perception adaptive adjustment method to obtain image point cloud enhancement features, including the following steps: environmental feature extraction, environmental perception scoring, and dynamic enhancement.

[0012] The environmental feature extraction specifically involves extracting light intensity, visibility, rain and snow index, humidity, image clarity, and point cloud noise level from multi-source sensor data, and then normalizing and splicing them into a multi-dimensional environmental state feature vector.

[0013] The environmental perception score is used to quantify the impact of the current environment on perception performance. Specifically, it is calculated by constructing a weighted environmental scoring function based on a multi-dimensional environmental state feature vector.

[0014] The dynamic enhancement specifically involves generating a perception adjustment matrix based on the comprehensive environmental perception score, and then using the perception adjustment matrix as a parameter control signal to dynamically adjust the perception enhancement strategy to enhance the image frame data and point cloud frame data in the multi-source sensor data, thereby obtaining image point cloud enhancement features.

[0015] The perception adjustment matrix includes a brightness enhancement coefficient, a contrast adjustment coefficient, a defogging intensity coefficient, a point cloud noise filtering radius, a point cloud density compensation threshold, and a feature point extraction weight coefficient.

[0016] The image point cloud enhancement feature includes enhanced image frame data and point cloud frame data;

[0017] The enhancement processing functions include image enhancement algorithms and point cloud enhancement algorithms.

[0018] Furthermore, the target recognition is used to detect targets and identify anomalies. Specifically, it employs an environment-aware-driven multi-level target recognition method to perform target recognition and obtain target recognition reference data, including the following steps: multi-level feature detection, feature adjustment, motion joint discrimination, abnormal target screening, and target recognition reference data generation.

[0019] The multi-level feature detection specifically involves dynamically adjusting the key parameters of the multi-level detector based on the perception adjustment matrix, constructing a multi-level detector to perform layered detection of the enhanced features of the image point cloud, outputting the detection results level by level, and deduplicating and merging the targets obtained from each level of detection to obtain a target set.

[0020] The multi-level detector adopts a multi-scale target detection architecture based on deep convolutional neural networks, including three levels: coarse detection, fine detection, and micro detection. The coarse detection is used to quickly detect large-scale targets, the fine detection is used to detect medium-scale targets, and the micro detection is used to detect tiny targets.

[0021] The feature adjustment specifically involves linearly mapping the perceptual adjustment matrix through a fully connected layer to generate an adjustment weight vector, and extracting the feature vector corresponding to each target from the target set to obtain the target feature vector; then, by multiplying the adjustment weight vector and the target feature vector element by element, a target adjustment feature vector set is generated.

[0022] The motion joint discrimination specifically involves obtaining a target reference feature vector from a pre-built target library to represent the standard appearance features of the corresponding target in an ideal environment, predicting the motion trajectory of the target set in two consecutive frames using a target tracking algorithm to generate a target motion feature vector, reconstructing the target motion feature vector using a pre-trained autoencoder, and calculating the reconstruction error as the target motion anomaly score; then combining the target reference feature vector, the target motion anomaly score, and the target adjustment feature vector, performing a weighted summation to calculate the comprehensive target anomaly score.

[0023] The abnormal target screening specifically involves comparing the comprehensive score of the target's abnormality with an abnormality threshold. When the comprehensive score of the target's abnormality is greater than the abnormality threshold, the corresponding target is marked as an abnormal target, and an abnormal target set is generated.

[0024] The target recognition reference data generation specifically involves performing the feature detection, feature adjustment, motion joint discrimination, and abnormal target screening to generate target recognition reference data, which includes an environmental perception comprehensive score, a target anomaly comprehensive score, and an abnormal target set.

[0025] Furthermore, the aforementioned risk warning specifically involves generating warning information based on target identification reference data and through comprehensive analysis of environmental risks and target anomalies, including the following steps: comprehensive anomaly scoring and warning reminder;

[0026] The anomaly comprehensive scoring is specifically based on target identification reference data to perform anomaly comprehensive scoring and generate a regional risk index;

[0027] Specifically, when the regional risk index exceeds the risk threshold, an early warning message is generated and sent. The early warning message includes the regional risk index, a set of abnormal targets, the warning location, and a timestamp.

[0028] The beneficial effects achieved by the present invention using the above solution are as follows:

[0029] (1) In response to the technical problems in the existing autonomous driving perception enhancement process, such as poor adaptability to complex environments and large fluctuations in perception image quality, which leads to a significant decrease in target recognition accuracy under severe weather conditions and easy omission of key targets, this solution creatively adopts an environmental perception adaptive adjustment method for adaptive perception enhancement, which perceives different inspection environments in real time and adjusts the image and point cloud quality as needed, thereby maintaining a stable and clear perception effect under complex weather conditions such as night, rain, snow, fog and haze, providing reliable data support for subsequent target recognition, and effectively meeting the police scenario's requirements for all-weather, high-accuracy, and low-misjudgment intelligent inspection;

[0030] (2) In response to the technical problem that the target recognition accuracy is greatly affected by environmental interference in the existing autonomous driving perception enhancement process, and small targets or fast-moving targets are difficult to identify, which affects the inspection effect, this solution creatively adopts a multi-level target recognition method driven by environmental perception to identify targets, comprehensively identify targets of different types, sizes and behavioral states, and can intelligently adjust the recognition strategy according to the environmental conditions. Even in complex environments, it can still stably identify key targets and abnormal situations, significantly improving the recognition reliability of police unmanned patrol vehicles in all weather and multiple scenarios.

[0031] (3) In response to the technical problems in the existing autonomous driving perception enhancement process, such as the untimely response to sudden environmental changes and suspicious target behavior, and the failure to comprehensively consider the environmental background and target behavior in the risk assessment process, resulting in incomplete early warning information, this solution creatively conducts risk early warning by comprehensively analyzing environmental risks and target anomalies, realizing real-time perception and intelligent response to potentially high-risk areas. It can not only detect abnormal behavior in complex environments in advance, but also dynamically assess risks in combination with the current environmental status, thereby effectively improving the risk identification capability and handling efficiency of police unmanned patrol vehicles in various environments. Attached Figure Description

[0032] Figure 1 A structural block diagram of an autonomous driving perception enhancement system for police unmanned patrol vehicles provided by the present invention;

[0033] Figure 2 This is a flowchart illustrating the adaptive perception enhancement module.

[0034] Figure 3 This is a flowchart of the target recognition module.

[0035] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0038] Example 1, see Figure 1 The present invention provides an autonomous driving perception enhancement system for police unmanned patrol vehicles, comprising: a sensor data acquisition module, an adaptive perception enhancement module, a target recognition module, and a risk warning module;

[0039] The sensor data acquisition module is used for sensor data acquisition. Through sensor data acquisition, multi-source sensor data is obtained, and the multi-source sensor data is sent to the adaptive perception enhancement module.

[0040] The adaptive perception enhancement module is used for adaptive perception enhancement. Based on multi-source sensor data, it adopts an environmental perception adaptive adjustment method to perform adaptive perception enhancement, obtains image point cloud enhancement features, and sends the image point cloud enhancement features to the target recognition module.

[0041] The target recognition module is used for target recognition. Based on the enhanced features of the image point cloud, it uses an environment-aware-driven multi-level target recognition method to perform target recognition, obtain target recognition reference data, and send the target recognition reference data to the risk warning module.

[0042] The risk warning module is used for risk warning. Based on target identification reference data, it conducts risk warning by comprehensively analyzing environmental risks and target anomalies, and generates warning information.

[0043] Example 2, see Figure 1 This embodiment is based on the above embodiment. Specifically, during the patrol process of the police unmanned patrol vehicle, image frame data, point cloud frame data, inertial measurement data and meteorological data are collected in real time through the vehicle-mounted multimodal sensor system. Time alignment is performed through a timestamp synchronization mechanism to obtain synchronized multi-source data. Then, according to the spatial calibration relationship between the sensors, the synchronized multi-source data is spatially aligned to obtain multi-source sensor data.

[0044] The meteorological data includes temperature, humidity, precipitation type, wind speed, wind direction, air pressure, and ambient light.

[0045] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. The adaptive perception enhancement is used to perceive the environmental state during the inspection process. The adaptive perception enhancement is performed by using an environmental perception adaptive adjustment method to obtain image point cloud enhancement features. It includes the following steps: environmental feature extraction, environmental perception scoring and dynamic enhancement.

[0046] The environmental feature extraction specifically involves extracting light intensity, visibility, rain and snow index, humidity, image clarity, and point cloud noise level from multi-source sensor data, and then normalizing and splicing them into a multi-dimensional environmental state feature vector.

[0047] The environmental perception score is used to quantify the impact of the current environment on perception performance. Specifically, it is calculated based on a multi-dimensional environmental state feature vector by constructing a weighted environmental scoring function. The calculation formula is as follows:

[0048] ;

[0049] ;

[0050] In the formula, S env This is the comprehensive environmental perception score, where i is the feature component index. f is the sensitivity coefficient of the i-th feature component. i (E i E is the transformation function of the i-th eigencomponent.i is the i-th feature component in the multidimensional environmental state feature vector, where L is the light intensity, V is the visibility, C is the image sharpness, W is the rain and snow index, H is the humidity, and P is the point cloud noise level.

[0051] The dynamic enhancement specifically involves generating a perception adjustment matrix based on the comprehensive environmental perception score, and then using the perception adjustment matrix as a parameter control signal to dynamically adjust the perception enhancement strategy to enhance image frame data and point cloud frame data in multi-source sensor data, thereby obtaining image and point cloud enhancement features. The calculation formula is as follows:

[0052] ;

[0053] ;

[0054] In the formula, M adj It is the perception adjustment matrix, used to provide parameter configuration for the enhancement processing function, M normal This is the normal mode parameter matrix, used to adapt to favorable environments. It is the first environmental threshold, M enhanced It is an enhanced mode parameter matrix used to adapt to medium-level environments. It is the second environmental threshold. M safe This is a safety mode parameter matrix used to adapt to harsh environments, F enh It is an image point cloud enhancement feature, f enh (·) is the enhancement processing function, I is the image frame data, and P is the point cloud frame data;

[0055] The perception adjustment matrix includes a brightness enhancement coefficient, a contrast adjustment coefficient, a defogging intensity coefficient, a point cloud noise filtering radius, a point cloud density compensation threshold, and a feature point extraction weight coefficient.

[0056] The image point cloud enhancement feature includes enhanced image frame data and point cloud frame data;

[0057] The enhancement processing function includes an image enhancement algorithm and a point cloud enhancement algorithm. The image enhancement algorithm includes brightness enhancement, contrast enhancement, and dehazing. The point cloud enhancement algorithm includes noise filtering, density compensation, and feature point extraction enhancement.

[0058] By performing the above operations, this solution addresses the technical problems in existing autonomous driving perception enhancement processes, such as poor adaptability to complex environments, large fluctuations in perceived image quality, leading to a significant decrease in target recognition accuracy and easy omission of key targets under adverse weather conditions. It creatively employs an environmental perception adaptive adjustment method for adaptive perception enhancement, perceiving different inspection environments in real time and adjusting image and point cloud quality as needed. This maintains stable and clear perception under complex weather conditions such as nighttime, rain, snow, and fog, providing reliable data support for subsequent target recognition and effectively meeting the police scenario's demand for all-weather, high-accuracy, and low-false-judgment intelligent inspection.

[0059] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. The target recognition is used to detect targets and identify anomalies. Specifically, it adopts an environment perception-driven multi-level target recognition method to perform target recognition and obtain target recognition reference data. It includes the following steps: multi-level feature detection, feature adjustment, motion joint discrimination, abnormal target screening and target recognition reference data generation.

[0060] The multi-level feature detection specifically involves dynamically adjusting the key parameters of the multi-level detector based on the perception adjustment matrix, constructing a multi-level detector to perform layered detection of the enhanced features of the image point cloud, outputting the detection results level by level, and deduplicating and merging the targets obtained from each level of detection to obtain a target set.

[0061] The calculation formula for dynamically adjusting the key parameters of the multi-level detector based on the sensing adjustment matrix is ​​as follows:

[0062] ;

[0063] ;

[0064] In the formula, It is the dynamically adjusted confidence threshold of the multi-level detector, and max(·) is the function to take the maximum value. This is the lower limit of the confidence threshold for multi-level detectors, used to prevent a surge in false detections caused by excessively lowering the confidence threshold of multi-level detectors. It is the base confidence threshold for multi-level detectors. It is the confidence level adjustment coefficient. It is the defogging intensity coefficient. It is the brightness enhancement factor. This is the dynamically adjusted NMS threshold for the multi-stage detector. `clip(·)` is the cutoff function used to ensure that the dynamically adjusted NMS threshold for the multi-stage detector is within [...]. , ] interval, It is the basic NMS threshold of the multi-level detector. It is the NMS adjustment coefficient, r filter It is the point cloud noise filtering radius. It is the maximum point cloud noise filtering radius. It is the lower limit of the NMS threshold for multi-level detectors. It is the upper limit of the NMS threshold for multi-level detectors;

[0065] The multi-level detector adopts a multi-scale target detection architecture based on deep convolutional neural networks, including three levels: coarse detection, fine detection, and micro detection. The coarse detection is used to quickly detect large-scale targets, the fine detection is used to detect medium-scale targets, and the micro detection is used to detect tiny targets.

[0066] The feature adjustment specifically involves linearly mapping the perceptual adjustment matrix through a fully connected layer to generate an adjustment weight vector, and extracting the feature vector corresponding to each target from the target set to obtain the target feature vector; then, by element-wise multiplying the adjustment weight vector and the target feature vector, a target adjusted feature vector set is generated, calculated using the following formula:

[0067] ;

[0068] ;

[0069] In the formula, W is the adjusted feature vector of the j-th target, where j is the target index. adj It is an adjustment weight vector. It is the element-wise multiplication symbol, fea j It is the j-th target feature vector, Fea adj It is the set of target-adjusted feature vectors, where N is the number of targets;

[0070] The motion joint discrimination specifically involves obtaining a target reference feature vector from a pre-built target library to represent the standard appearance features of the corresponding target in an ideal environment. A target tracking algorithm is then used to predict the motion trajectory of the target set across two consecutive frames, generating a target motion feature vector. This vector is then reconstructed using a pre-trained autoencoder, and the reconstruction error is calculated as the target motion anomaly score. Finally, the target reference feature vector, the target motion anomaly score, and the target adjustment feature vector are combined and weighted to calculate a comprehensive target anomaly score. The calculation formula is as follows:

[0071] ;

[0072] In the formula, C j It is the comprehensive score of the j-th target anomaly. It is a static similarity weight, Sim cos (·) is the cosine similarity function, fea refIt is the target reference feature vector. It is the weight of motion anomalies, s j It is the motion anomaly score of the j-th target;

[0073] The pre-built target library is specifically constructed by collecting multi-source sensor data of typical targets under good conditions, extracting static feature vectors of typical targets using standard image processing and point cloud modeling algorithms, and combining manual annotation to annotate target categories and bounding boxes, thereby constructing a target reference feature database as the pre-built target library.

[0074] The typical targets include pedestrians, motor vehicles, non-motorized vehicles, road obstacles, construction facilities, and temporary stockpiles;

[0075] The target motion feature vector includes the target's motion velocity, acceleration, and trajectory stability;

[0076] The pre-trained autoencoder uses target motion feature vectors with normal trajectories for unsupervised training to learn the motion behavior patterns of normal targets, thereby generating a higher reconstruction error when reconstructing abnormal targets.

[0077] The abnormal target screening specifically involves comparing the comprehensive score of the target's abnormality with an abnormality threshold. When the comprehensive score of the target's abnormality is greater than the abnormality threshold, the corresponding target is marked as an abnormal target, and an abnormal target set is generated.

[0078] The target recognition reference data generation specifically involves generating target recognition reference data by performing the feature detection, feature adjustment, motion joint discrimination, and abnormal target screening. The target recognition reference data includes an environmental perception comprehensive score, a target anomaly comprehensive score, and an abnormal target set.

[0079] By performing the above operations, this solution addresses the technical problems in existing autonomous driving perception enhancement processes, such as the target recognition accuracy being greatly affected by environmental interference and the difficulty in recognizing small or fast-moving targets, which affects the inspection effect. This solution creatively adopts a multi-level target recognition method driven by environmental perception to comprehensively identify targets of different types, sizes, and behavioral states. It can also intelligently adjust the recognition strategy according to environmental conditions, and can still stably identify key targets and abnormal situations in complex environments, significantly improving the recognition reliability of police unmanned patrol vehicles in all weather and multi-scenario environments.

[0080] Example 5, see Figure 1 This embodiment is based on the above embodiment. The risk warning is specifically based on target identification reference data, and generates warning information by comprehensively analyzing environmental risks and target anomalies. It includes the following steps: anomaly comprehensive scoring and warning reminder.

[0081] The aforementioned anomaly comprehensive scoring is specifically based on target identification reference data to perform anomaly comprehensive scoring and generate a regional risk index. The calculation formula is as follows:

[0082] ;

[0083] In the formula, R is the regional risk index. It is the environmental anomaly weight. It is the target anomaly weight. It is the maximum value of the overall target anomaly score among all targets;

[0084] Specifically, the early warning notification generates and sends an early warning message when the regional risk index exceeds the risk threshold. The early warning message includes the regional risk index, the set of abnormal targets, the warning location, and a timestamp.

[0085] By performing the above operations, this solution addresses the technical problems in existing autonomous driving perception enhancement processes, such as untimely response to sudden environmental changes and suspicious target behavior, and incomplete early warning information due to the failure to comprehensively consider environmental background and target behavior during risk assessment. This solution creatively conducts risk warnings by comprehensively analyzing environmental risks and target anomalies, achieving real-time perception and intelligent response to potentially high-risk areas. It can not only detect abnormal behavior in complex environments in advance, but also dynamically assess risks based on the current environmental state, thereby effectively improving the risk identification capabilities and handling efficiency of police unmanned patrol vehicles in various environments.

[0086] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0087] 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.

[0088] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An autonomous driving perception enhancement system for police unmanned patrol vehicles, characterized in that: It includes a sensor data acquisition module, an adaptive perception enhancement module, a target recognition module, and a risk warning module; The sensor data acquisition module is used for sensor data acquisition. Through sensor data acquisition, multi-source sensor data is obtained, and the multi-source sensor data is sent to the adaptive perception enhancement module. The adaptive perception enhancement module is used for adaptive perception enhancement. Based on multi-source sensor data, it adopts an environmental perception adaptive adjustment method to perform adaptive perception enhancement, obtains image point cloud enhancement features, and sends the image point cloud enhancement features to the target recognition module. The target recognition module is used for target recognition. Based on the enhanced features of the image point cloud, it uses an environment-aware-driven multi-level target recognition method to perform target recognition, obtain target recognition reference data, and send the target recognition reference data to the risk warning module. The risk warning module is used for risk warning. Based on target identification reference data, it conducts risk warning by comprehensively analyzing environmental risks and target anomalies, and generates warning information.

2. The autonomous driving perception enhancement system for police unmanned patrol vehicles according to claim 1, characterized in that: The adaptive perception enhancement is used to perceive the environmental state during the inspection process. It adopts an environmental perception adaptive adjustment method to perform adaptive perception enhancement and obtain image point cloud enhancement features. It includes the following steps: environmental feature extraction, environmental perception scoring and dynamic enhancement. The environmental feature extraction specifically involves extracting light intensity, visibility, rain and snow index, humidity, image clarity, and point cloud noise level from multi-source sensor data, and then normalizing and splicing them into a multi-dimensional environmental state feature vector. The environmental perception score is used to quantify the impact of the current environment on perception performance. Specifically, it is calculated by constructing a weighted environmental scoring function based on a multi-dimensional environmental state feature vector. The dynamic enhancement specifically involves generating a perception adjustment matrix based on the comprehensive environmental perception score, and then using the perception adjustment matrix as a parameter control signal to dynamically adjust the perception enhancement strategy to enhance the image frame data and point cloud frame data in the multi-source sensor data, thereby obtaining image point cloud enhancement features.

3. The autonomous driving perception enhancement system for police unmanned patrol vehicles according to claim 2, characterized in that: In the dynamic enhancement, the perceptual adjustment matrix includes a brightness enhancement coefficient, a contrast adjustment coefficient, a dehazing intensity coefficient, a point cloud noise filtering radius, a point cloud density compensation threshold, and a feature point extraction weight coefficient; the image point cloud enhancement features include enhanced image frame data and point cloud frame data; and the enhancement processing function includes an image enhancement algorithm and a point cloud enhancement algorithm.

4. The autonomous driving perception enhancement system for police unmanned patrol vehicles according to claim 3, characterized in that: The target recognition is used to detect targets and identify anomalies. Specifically, it adopts a multi-level target recognition method driven by environmental perception to perform target recognition and obtain target recognition reference data. It includes the following steps: multi-level feature detection, feature adjustment, joint motion discrimination, abnormal target screening and target recognition reference data generation. The multi-level feature detection specifically involves dynamically adjusting the key parameters of the multi-level detector based on the perception adjustment matrix, constructing a multi-level detector to perform layered detection of the enhanced features of the image point cloud, outputting the detection results level by level, and deduplicating and merging the targets obtained from each level of detection to obtain a target set. The feature adjustment specifically involves linearly mapping the perceptual adjustment matrix through a fully connected layer to generate an adjustment weight vector, and extracting the feature vector corresponding to each target from the target set to obtain the target feature vector; then, by multiplying the adjustment weight vector and the target feature vector element by element, a target adjustment feature vector set is generated. The motion joint discrimination specifically involves obtaining a target reference feature vector from a pre-built target library to represent the standard appearance features of the corresponding target in an ideal environment, predicting the motion trajectory of the target set in two consecutive frames using a target tracking algorithm to generate a target motion feature vector, reconstructing the target motion feature vector using a pre-trained autoencoder, and calculating the reconstruction error as the target motion anomaly score; then combining the target reference feature vector, the target motion anomaly score, and the target adjustment feature vector, performing a weighted summation to calculate the comprehensive target anomaly score. The abnormal target screening specifically involves comparing the comprehensive score of the target's abnormality with an abnormality threshold. When the comprehensive score of the target's abnormality is greater than the abnormality threshold, the corresponding target is marked as an abnormal target, and an abnormal target set is generated. The target recognition reference data generation specifically involves performing the feature detection, feature adjustment, motion joint discrimination, and abnormal target screening to generate target recognition reference data, which includes an environmental perception comprehensive score, a target anomaly comprehensive score, and an abnormal target set.

5. The autonomous driving perception enhancement system for police unmanned patrol vehicles according to claim 4, characterized in that: In the multi-level detection, the multi-level detector adopts a multi-scale target detection architecture based on deep convolutional neural networks, including three levels: coarse detection, fine detection, and micro detection. The coarse detection is used to quickly detect large-scale targets, the fine detection is used to detect medium-scale targets, and the micro detection is used to detect tiny targets.

6. The autonomous driving perception enhancement system for police unmanned patrol vehicles according to claim 5, characterized in that: The risk warning is specifically based on target identification reference data, and generates warning information by comprehensively analyzing environmental risks and target anomalies. It includes the following steps: comprehensive anomaly scoring and warning reminder. The anomaly comprehensive scoring is specifically based on target identification reference data to perform anomaly comprehensive scoring and generate a regional risk index; Specifically, when the regional risk index exceeds the risk threshold, an early warning message is generated and sent. The early warning message includes the regional risk index, a set of abnormal targets, the warning location, and a timestamp.

7. The autonomous driving perception enhancement system for police unmanned patrol vehicles according to claim 6, characterized in that: The sensor data acquisition specifically involves collecting image frame data, point cloud frame data, inertial measurement data, and meteorological data in real time through an onboard multimodal sensor system during the patrol process of the police unmanned patrol vehicle. Time alignment is performed through a timestamp synchronization mechanism to obtain synchronized multi-source data. Then, based on the spatial calibration relationship between the sensors, the synchronized multi-source data is spatially aligned to obtain multi-source sensor data.

Images