Dynamic focal point target detection method and system for autonomous vehicle

Abstract

The application discloses a dynamic focus target detection method and system for an automatic driving vehicle, relates to the technical field of automatic driving, and discloses a dynamic focus target detection method and system for an automatic driving vehicle, which obtains vehicle images and speed data, generates an image motion vector field based on motion analysis, performs dynamic region priority division in combination with vehicle speed, outputs dynamic detection frequency planning data, generates a to-be-processed region mask by using an adaptive sparse decision module, selectively crops an image input target detection neural network, and finally maps coordinates to output results, so that intelligent allocation of computing resources is realized, computing resources can be efficiently allocated, processing delay can be reduced, and the accuracy and real-time performance of key target detection can be improved.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to a method and system for dynamic focus target detection for autonomous vehicles. Background Technology

[0002] In the actual operation of autonomous driving systems, real-time object detection is a core component for ensuring driving safety, and its performance directly determines the vehicle's perception and response speed of the surrounding environment. Current mainstream technologies generally rely on high-performance computing hardware to perform intensive processing of each frame continuously captured by onboard cameras across the entire image. However, with the proliferation of complex scenarios such as urban roads and highways, image data exhibits significant non-uniformity: a large number of pixels are concentrated in static background areas, such as fixed buildings, road markings, and the sky. These areas change only slightly or almost constantly over time, yet consume the majority of the processor's computing resources. Meanwhile, critical dynamic targets such as pedestrians, non-motorized vehicles, or vehicles suddenly entering the scene often appear abruptly in specific locations within the field of view, with short durations and unpredictable locations. The inherent flaw of the full-frame detection mode is its inability to effectively identify and prioritize high-risk areas, leading to excessive consumption of computing resources in information-redundant areas and a significant increase in overall system processing latency. This latency is particularly pronounced at high speeds, potentially causing the system to miss millisecond-level critical response windows, directly threatening driving safety. While existing research has attempted to introduce motion detection mechanisms or pre-defined fixed-region sampling strategies to reduce computational load, these methods have significant limitations. Motion detection schemes typically filter regions based solely on local pixel changes, failing to integrate the vehicle's real-time motion state (such as speed) for global coordination. For example, in high-speed scenarios, they cannot automatically prioritize the detection of areas ahead. Fixed-region sampling, on the other hand, lacks adaptability to dynamic scene changes, potentially over-detecting static areas at low speeds and easily missing sudden targets in edge areas at high speeds. Furthermore, existing technologies struggle to balance detection accuracy with resource consumption, and under complex lighting or occlusion conditions, critical targets may still be missed or falsely detected.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this application is to provide a dynamic focus target detection method and system for autonomous vehicles, which aims to efficiently allocate computing resources, reduce processing latency, and improve the accuracy and real-time performance of key target detection.

[0005] To achieve the above objectives, this application proposes a dynamic focus target detection method for autonomous vehicles, the method comprising: Acquire current image frames captured by the autonomous vehicle during its operation, as well as vehicle speed data from the vehicle bus interface; Motion analysis is performed based on the current image frame and at least one historical image frame to output an image motion vector field that reflects the direction and magnitude of pixel motion. Using the image motion vector field and the vehicle speed data as input, dynamic region priority partitioning is performed, and dynamic detection frequency planning data is output. The dynamic detection frequency planning data is used to define a two-dimensional Boolean decision matrix for whether different spatial regions in the current image frame are processed in this detection. The current image frame, the dynamic detection frequency planning data, and the historical image frames are used as inputs to the adaptive sparse decision module for processing, and the output is a mask of the unprocessed region with the same size as the current image frame; the mask of the unprocessed region is used as a binary matrix to identify the pixel region to be detected. Using the mask of the region to be processed and the current image frame as input, selective region cropping and stitching are performed, and the target candidate region image is output, which is composed of the image content of the region to be detected by the mask. The target candidate region image is input into the target detection neural network for processing, and an initial target detection result is output; the initial target detection result includes the category of at least one target and its bounding box coordinates in the target candidate region image; Using the initial target detection result and the spatial location mapping relationship recorded by the mask of the region to be processed as input, coordinate inverse mapping processing is performed to map the target bounding box coordinates to the original coordinate system of the current image frame, and the final target detection result is output.

[0006] In one embodiment, the step of using the image motion vector field and the vehicle speed data as input to perform dynamic region priority partitioning processing and output dynamic detection frequency planning data includes: The current image frame is divided into multiple image sub-regions, and the image motion vector field is used as input to calculate and output the motion change intensity value corresponding to each image sub-region. Using the vehicle speed data as input, a preset vehicle speed-frequency mapping table is queried to determine and output the global detection frequency reference value; Using the motion change intensity value and the global detection frequency reference value as input, local modulation processing is performed to calculate and output the corresponding dynamic detection frequency value for each image sub-region. The dynamic detection frequency value of each image sub-region is compared with a preset frequency threshold to generate and output the dynamic detection frequency planning data.

[0007] In one embodiment, the step of dividing the current image frame into multiple image sub-regions and using the image motion vector field as input to calculate and output the motion change intensity value corresponding to each image sub-region includes: For each image sub-region, extract all image motion vectors that fall within its boundary range; Calculate the amplitude of each extracted image motion vector to obtain the set of amplitude values ​​for that sub-region; Statistical analysis is performed on the set of amplitude values ​​to obtain the current frame statistical characteristics of the set; the statistical characteristics include at least one of the average amplitude, maximum amplitude, and standard deviation of amplitude. Read the historical motion change intensity value corresponding to the sub-region of the image from multiple consecutive historical image frames; The statistical features of the current frame are weighted and fused with the historical motion change intensity values, and the fusion result is output as the final motion change intensity value of the image sub-region.

[0008] In one embodiment, the step of using the vehicle speed data as input, querying a preset vehicle speed-frequency mapping table, and determining and outputting a global detection frequency reference value includes: Determine the vehicle speed range to which the vehicle speed data belongs; The reference frequency value corresponding to the vehicle speed range is retrieved from the preset vehicle speed-frequency mapping table, and the reference frequency value is used as the output global detection frequency reference value; wherein, the preset vehicle speed-frequency mapping table contains multiple discrete vehicle speed ranges and their corresponding reference frequency values.

[0009] In one embodiment, the step of using the motion change intensity value and the global detection frequency reference value as input, performing local modulation processing, and calculating and outputting the corresponding dynamic detection frequency value for each image sub-region includes: Based on the motion change intensity value of each image sub-region, the preset motion intensity-modulation coefficient mapping relationship is queried to obtain and output the corresponding local modulation coefficient; The global detection frequency reference value is multiplied by the local modulation coefficient to obtain and output the preliminary frequency value of the image sub-region; An upper limit constraint and a lower limit constraint are applied to the initial frequency value, and the constrained value is output as the dynamic detection frequency value of the image sub-region; wherein, in the preset motion intensity-modulation coefficient mapping relationship, the motion change intensity value is proportional to the corresponding modulation coefficient.

[0010] In one embodiment, the step of taking the current image frame, the dynamic detection frequency planning data, and the historical image frames as inputs, feeding them into an adaptive sparse decision module for processing, and outputting a mask for the region to be processed that has the same size as the current image frame includes: The dynamic detection frequency planning data is analyzed to identify all image sub-regions that are planned to be detected in the current frame, and the corresponding spatial positions in the mask of the region to be processed are marked as the first identifier value. For the image sub-regions in the dynamic detection frequency planning data that are planned not to be detected in the current frame, perform spatiotemporal content consistency verification. The first identifier value and the updated identifier value after verification are integrated to form and output the complete mask of the region to be processed.

[0011] In one embodiment, the step of performing spatiotemporal content consistency verification for image sub-regions planned as not requiring detection in the current frame in the dynamic detection frequency planning data includes: Extract the current image block of the image sub-region that does not need to be detected from the current image frame, and extract the historical image block sequence corresponding to the image sub-region from multiple historical image frames; Calculate the content difference between the current image block and each image block in the historical image block sequence, and aggregate and output a comprehensive content difference value; If the overall content difference value exceeds the preset consistency threshold, the position of the image sub-region in the mask of the region to be processed will be updated and marked with the first identifier value. If the overall content difference value does not exceed the consistency threshold, then its corresponding position in the mask of the area to be processed is marked as the second identifier value.

[0012] In one embodiment, the step of calculating the content difference between the current image patch and each image patch in the historical image patch sequence, and aggregating and outputting a comprehensive content difference value, includes: Feature extraction is performed on the current image patch, and the current feature vector is output; Feature extraction is performed on each historical image block in the historical image block sequence, and multiple historical feature vectors are output accordingly. Calculate the distance between the current feature vector and each of the historical feature vectors respectively to obtain multiple corresponding distance values; The average value of the multiple distance values ​​is calculated and the average value is output as the comprehensive content difference value.

[0013] In one embodiment, the method further includes: By comparing the dynamic detection frequency planning data with the mask of the region to be processed, all image sub-regions that are planned as not needing detection in the dynamic detection frequency planning data but are marked as needing detection in the mask of the region to be processed are identified, thus obtaining a set of difference regions. Spatiotemporal distribution statistics are performed on the set of differential regions to calculate their aggregation intensity in image frames and their frequency of occurrence as a function of vehicle speed. Based on the aggregation intensity and the occurrence frequency, determine whether the decision logic for generating the two-dimensional Boolean decision matrix needs optimization; if it is determined that optimization is needed, generate corresponding optimization instructions to adjust the decision logic.

[0014] Furthermore, to achieve the above objectives, this application also proposes a dynamic focus target detection system for autonomous vehicles, the dynamic focus target detection system for autonomous vehicles comprising: a memory, a processor, and a dynamic focus target detection program for autonomous vehicles stored in the memory and executable on the processor, the dynamic focus target detection program for autonomous vehicles being configured to implement the steps of the dynamic focus target detection method for autonomous vehicles.

[0015] The proposed method and system for dynamic focus target detection in autonomous vehicles acquires vehicle images and speed data, generates image motion vector fields based on motion analysis, performs dynamic region priority division based on vehicle speed, outputs dynamic detection frequency planning data, generates a mask for the region to be processed using an adaptive sparse decision module, selectively crops the image and inputs it into the target detection neural network, and finally outputs the results by mapping coordinates. This achieves intelligent allocation of computing resources, which can efficiently allocate computing resources, reduce processing latency, and improve the accuracy and real-time performance of key target detection. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an embodiment of the dynamic focus target detection method for autonomous vehicles according to this application. Figure 2This is a schematic diagram of a structure provided for an embodiment of the dynamic focus target detection system for autonomous vehicles according to this application.

[0019] Explanation of icon numbers: 10. Memory; 20. Processor.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of this application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] It should be understood that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] In the field of autonomous driving, existing target detection methods typically perform full-image detection for each frame, leading to wasted computational resources, increased processing latency, and difficulty in meeting the real-time response requirements of highly dynamic scenes. Existing motion detection or fixed-region sampling methods lack collaborative perception of the global state of the environment and local motion changes, and cannot adaptively and dynamically allocate detection resources, resulting in the risk of large response latency or missed detection of key targets.

[0024] Based on this, embodiments of this application provide a dynamic focus target detection method for autonomous vehicles, referring to... Figure 1 The dynamic focus target detection method for autonomous vehicles includes steps S100 to S700, wherein: Step S100: Obtain the current image frame collected by the autonomous vehicle during driving, as well as the vehicle speed data from the vehicle bus interface; Step S200: Motion analysis is performed based on the current image frame and at least one historical image frame to output an image motion vector field that reflects the direction and magnitude of pixel motion. Step S300: Using the image motion vector field and the vehicle speed data as input, perform dynamic region priority division processing and output dynamic detection frequency planning data; the dynamic detection frequency planning data is used to define a two-dimensional Boolean decision matrix for whether different spatial regions in the current image frame are processed in this detection. Step S400: The current image frame, the dynamic detection frequency planning data, and the historical image frames are used as inputs to the adaptive sparse decision module for processing, and the output is a mask of the region to be processed with the same size as the current image frame; the mask of the region to be processed is a binary matrix used to identify the pixel region to be detected. Step S500: Using the mask of the region to be processed and the current image frame as input, perform selective region cropping and stitching processing, and output a target candidate region image composed of image content identified by the mask as the region to be detected; Step S600: Input the target candidate region image into the target detection neural network for processing and output an initial target detection result; the initial target detection result includes the category of at least one target and its bounding box coordinates in the target candidate region image; Step S700: Using the initial target detection result and the spatial position mapping relationship recorded by the mask of the region to be processed as input, perform coordinate reverse mapping processing to map the target bounding box coordinates to the original coordinate system of the current image frame, and output the final target detection result.

[0025] In this embodiment, the current image frame refers to a single frame image in the real-time video stream captured by the onboard camera at a specific moment by the autonomous vehicle. This image frame is the basic data for target detection analysis. Vehicle speed data refers to the current driving speed information of the autonomous vehicle obtained through the vehicle bus interface. This data reflects the overall motion state of the vehicle and can be used to assist in judging the relative motion of targets in the image. Historical image frames refer to multiple consecutive frames of images captured and stored by the onboard camera before the current image frame. These historical image frames, combined with the current image frame, can be used for motion analysis to identify the motion trend of pixels in the image. The image motion vector field refers to the motion direction and amplitude information calculated for each pixel or pixel block in the image after performing motion analysis on consecutive image frames. This vector field intuitively represents the changes in image content over time and is an important basis for identifying dynamic regions.

[0026] In this embodiment, the dynamic detection frequency planning data is specifically represented as a two-dimensional Boolean decision matrix. This matrix corresponds to the spatial layout of the current image frame and is used to indicate whether different spatial regions in the current image frame need to be processed in this target detection process. For example, a "true" value in the matrix indicates that the region needs to be detected, while a "false" value indicates that it does not need to be detected. The mask for the region to be processed is specifically represented as a binary matrix. This mask has the same size as the current image frame and is used to accurately identify the pixel regions in the current frame that need to be detected. One value in the mask (e.g., 1) indicates that the pixel region needs to be detected, and another value (e.g., 0) indicates that it does not need to be detected. The target candidate region image refers to the image content cropped and stitched from the current image frame according to the mask for the region to be processed. This image only contains the regions identified by the mask as to be detected, thereby reducing the processing range of the subsequent target detection neural network.

[0027] In this embodiment, the object detection neural network refers to a trained deep learning model whose function is to receive image input, identify specific targets present in the image, and output the target's category and its precise location in the image (usually represented by bounding box coordinates). The initial object detection result refers to the direct output of the object detection neural network after processing the target candidate region image. This result contains the category information of all targets detected in the target candidate region image and their corresponding bounding box coordinates. The final object detection result refers to the result of transforming the target bounding box coordinates from the initial object detection result back to the original coordinate system of the current image frame through a spatial location mapping relationship. This result is the final target information used by the autonomous driving system for decision-making and control.

[0028] In this embodiment, the dynamic focus target detection method for autonomous vehicles first acquires the current image frames collected during the autonomous vehicle's operation, as well as vehicle speed data from the vehicle bus interface. The current image frames can be captured in real-time by an onboard camera, for example, the camera continuously captures images at a frame rate of 30 frames per second or higher. Vehicle speed data can be obtained through the vehicle's internal sensor network or CAN bus, for example, by reading speed information provided by the vehicle's wheel speed sensors or GPS module. This data forms the basis for subsequent processing.

[0029] In this embodiment, next, motion analysis is performed based on the acquired current image frame and at least one historical image frame to output an image motion vector field reflecting the direction and magnitude of pixel motion. Motion analysis can be implemented in various ways; for example, optical flow can be used to calculate the trajectory and velocity of pixels in the image, or inter-frame differencing can be used to identify changing regions in the image. Thus, a vector field containing motion information of each region in the image can be obtained.

[0030] Furthermore, using the image motion vector field and vehicle speed data as input, dynamic region priority partitioning is performed, outputting dynamic detection frequency planning data. This dynamic detection frequency planning data is used to define a two-dimensional Boolean decision matrix for whether different spatial regions in the current image frame are processed in this detection. For example, based on regions with large motion amplitudes in the image motion vector field, combined with vehicle speed data, a preliminary judgment can be made as to which regions may contain dynamic targets. A simple implementation is to divide the image into several grids, roughly count the motion intensity within each grid, and manually set a threshold based on the vehicle speed. Grids above the threshold are marked as needing detection, while grids below the threshold are marked as not needing detection.

[0031] Based on this, the current image frame, the dynamic detection frequency planning data, and the historical image frames are used as inputs to the adaptive sparse decision module for processing, outputting a mask of the region to be processed with the same size as the current image frame. This mask is a binary matrix used to identify the pixel regions to be detected. For example, the adaptive sparse decision module can directly mark regions planned to be detected as to be detected in the mask based on the dynamic detection frequency planning data. For regions planned not to be detected, they can simply be marked as not to be detected in the mask.

[0032] In this embodiment, the system then uses the mask of the region to be processed and the current image frame as input to perform selective region cropping and stitching, outputting a target candidate region image composed of image content identified by the mask as the region to be detected. Specifically, the system iterates through the mask of the region to be processed, identifying all pixel regions marked as to be detected. Then, it extracts the corresponding pixel content from the current image frame and stitches these extracted image blocks together to form one or more target candidate region images. This processing method can effectively reduce the amount of input data for subsequent object detection.

[0033] In this embodiment, the target candidate region image is then input into the target detection neural network for processing, outputting an initial target detection result. This initial target detection result includes the category of at least one target and its bounding box coordinates within the target candidate region image. The target detection neural network can be any known deep learning model; for example, it can employ a two-stage detector based on region proposal or a single-stage detector. The neural network analyzes the input target candidate region image, identifies the target objects within it, and provides their category and precise location within the target candidate region image.

[0034] In this embodiment, the initial target detection result and the spatial location mapping relationship recorded by the mask of the region to be processed are used as input to perform coordinate inversion mapping processing. This maps the target bounding box coordinates to the original coordinate system of the current image frame, outputting the final target detection result. Since the initial target detection result is obtained on the cropped and stitched target candidate region image, its coordinate system differs from the original image frame. Therefore, it is necessary to utilize the original position information recorded by the mask of the region to be processed during the cropping and stitching process to accurately transform the bounding box coordinates of the detected target back to the original coordinate system of the current image frame. This allows for the accurate identification of the target detection result in the original image frame, which can then be used by the autonomous driving system.

[0035] In this embodiment, by dynamically acquiring vehicle speed data and image motion information, the image detection region is adaptively planned, and combined with spatiotemporal consistency verification, an accurate mask for the region to be processed is generated. Therefore, the target detection neural network only needs to process the critical dynamic regions, significantly reducing computational resource consumption and processing latency. This method, while ensuring high detection accuracy, effectively improves the real-time response capability of autonomous driving systems in high-dynamic scenarios such as urban roads or highway driving, and reduces the risk of missing key targets.

[0036] In one feasible implementation, the step of performing dynamic region priority partitioning processing and outputting dynamic detection frequency planning data using the image motion vector field and the vehicle speed data as input includes: dividing the current image frame into multiple image sub-regions, and using the image motion vector field as input to calculate and output the motion change intensity value corresponding to each image sub-region; using the vehicle speed data as input to query a preset vehicle speed-frequency mapping table, determining and outputting a global detection frequency reference value; using the motion change intensity value and the global detection frequency reference value as input to perform local modulation processing, calculating and outputting the corresponding dynamic detection frequency value for each image sub-region; comparing the dynamic detection frequency value of each image sub-region with a preset frequency threshold to generate and output the dynamic detection frequency planning data.

[0037] In this embodiment, the current image frame is first segmented into multiple image sub-regions, for example, using a fixed-size grid or adaptive segmentation based on image content. For each image sub-region, the system uses a previously generated image motion vector field as input. The image motion vector field reflects the direction and amplitude of pixel motion. By analyzing the image motion vectors falling within each sub-region, the motion change intensity value of that sub-region can be calculated. This value quantifies the activity level of motion within that sub-region, and can be represented, for example, by statistically analyzing the average, maximum, or standard deviation of the amplitudes of all motion vectors within that sub-region. Furthermore, to provide a more comprehensive implementation, this motion change intensity value can also be determined by analyzing the density, directional consistency, or relative motion with the background of the motion vectors within the sub-region.

[0038] In this embodiment, vehicle speed data is a crucial indicator of the autonomous vehicle's driving status. The system uses this vehicle speed data as input to query a pre-defined vehicle speed-frequency mapping table. This mapping table stores the relationship between different vehicle speed ranges and their corresponding global detection frequency benchmark values. For example, when the vehicle speed is low, the global detection frequency benchmark value may be low to conserve computing resources; when the vehicle speed is high, the global detection frequency benchmark value will increase accordingly to ensure timely response to rapidly changing environments. Besides discrete range mappings, this mapping table can also be defined using continuous functions, allowing the global detection frequency benchmark value to change smoothly with vehicle speed.

[0039] In this embodiment, after obtaining the motion change intensity value and the global detection frequency reference value for each image sub-region, the system performs local modulation processing. This processing aims to finely adjust the global reference frequency based on the actual motion activity of each sub-region. Specifically, a local modulation coefficient can be obtained by looking up a table or calculating a function based on the motion change intensity value. This modulation coefficient can be multiplied or added with the global detection frequency reference value to obtain the preliminary dynamic detection frequency value for that sub-region. To ensure the rationality of the detection frequency, the preliminary frequency value is also subject to upper and lower limits to avoid excessively high or low detection frequencies. The final output value after constraints is the dynamic detection frequency value for that image sub-region.

[0040] In this embodiment, the system calculates the dynamic detection frequency value for each image sub-region and compares it with a preset frequency threshold. If the dynamic detection frequency value of a sub-region is higher than or equal to the frequency threshold, it indicates that the sub-region needs to undergo target detection processing in the current frame; conversely, if it is lower than the threshold, it may indicate that the sub-region does not need to be detected in the current frame, or its detection priority can be reduced. Through this comparison, the system can generate and output dynamic detection frequency planning data, which is represented in the form of a two-dimensional Boolean decision matrix, clearly indicating which spatial regions in the current image frame will be processed in this detection, thereby achieving intelligent allocation of detection resources.

[0041] In this embodiment, through the above technical solution, this application can intelligently plan the frequency and region of target detection based on the real-time driving state of the autonomous vehicle and the local motion within the image frame. Specifically, the current image frame is divided into multiple image sub-regions, and the motion change intensity value of each sub-region is calculated, enabling the system to perceive the dynamic activity of the local scene. Simultaneously, a global detection frequency benchmark value is determined by combining vehicle speed data, ensuring the matching of the detection strategy with the overall driving environment of the vehicle. Based on this, through local modulation processing, a detection frequency matching its dynamic characteristics is calculated for each image sub-region, thereby achieving fine-grained allocation of detection resources. Finally, dynamic detection frequency planning data is generated by comparing with a frequency threshold, enabling the target detection system to concentrate computational resources on areas with active motion or those more important to driving safety, effectively avoiding repeated detection of static or low-priority areas, significantly improving the efficiency and real-time performance of target detection, while ensuring timely discovery of key targets.

[0042] In one feasible implementation, the step of dividing the current image frame into multiple image sub-regions and calculating and outputting the motion change intensity value corresponding to each image sub-region using the image motion vector field as input includes: for each image sub-region, extracting all image motion vectors falling within its boundary range; calculating the amplitude of each extracted image motion vector to obtain the amplitude value set of the sub-region; performing statistical analysis on the amplitude value set to obtain the current frame statistical features of the set; the statistical features include at least one of the amplitude average, amplitude maximum, and amplitude standard deviation; reading the historical motion change intensity value corresponding to the image sub-region from multiple consecutive historical image frames; weightedly fusing the current frame statistical features with the historical motion change intensity value, and outputting the fusion result as the final motion change intensity value of the image sub-region.

[0043] In this embodiment, when calculating the motion change intensity value of each image sub-region, firstly, for each image sub-region, the system extracts all image motion vectors falling within the boundary range of that image sub-region from the image motion vector field output by motion analysis of the current image frame and at least one historical image frame. These image motion vectors reflect the motion direction and amplitude of pixels within that sub-region. The image sub-region can be a preset fixed-size grid (e.g., 16x16 pixels) or an area adaptively divided according to image content or scene features.

[0044] In this embodiment, the system then calculates the amplitude of each extracted image motion vector. An image motion vector typically consists of a horizontal component and a vertical component, and its amplitude can be calculated using the Euclidean distance formula. This amplitude value represents the actual displacement distance of a pixel. By calculating the amplitudes of all extracted image motion vectors, a set of amplitude values ​​for the image sub-region in the current frame can be obtained. Subsequently, statistical analysis is performed on this set of amplitude values ​​to obtain the statistical characteristics of the set for the current frame. The statistical characteristics may include at least one of the following: average amplitude, maximum amplitude, and standard deviation of amplitude. The average amplitude reflects the overall motion intensity within the sub-region; the maximum amplitude indicates the presence of locally rapidly moving objects or significant motion events; and the standard deviation of amplitude reflects the consistency or dispersion of motion amplitude within the sub-region. For example, a larger standard deviation may indicate the presence of multiple motion patterns or objects at different speeds within the region. These statistical characteristics provide a comprehensive quantitative description of the motion state of the current frame.

[0045] In this embodiment, to introduce stability in the time dimension, the system reads the historical motion change intensity values ​​corresponding to the sub-region of the image from multiple consecutive historical image frames. These historical values ​​are typically stored in a circular buffer or a time series database, recording the evolution of the motion intensity of the sub-region over a past period.

[0046] In this embodiment, the current frame statistical features and the historical motion change intensity values ​​are finally weighted and fused. This weighted fusion aims to combine the real-time information of the current frame with the stability of historical data, thereby outputting a more robust and accurate motion change intensity value. For example, the fusion can be performed using a linear weighting method, i.e., final motion change intensity value = w1 × current frame statistical features + w2 × historical motion change intensity value, where w1 and w2 are weighting coefficients that can be adjusted according to the actual application scenario. The historical motion change intensity value can also be an exponentially weighted moving average of values ​​from multiple past frames. The fused result will be used as the final motion change intensity value for the image sub-region.

[0047] In this embodiment, by performing detailed amplitude statistical analysis on the image motion vectors within image sub-regions, this application can capture the motion features of the current frame more comprehensively and meticulously, avoiding information loss that may result from a single indicator. More importantly, by introducing motion change intensity values ​​from multiple consecutive historical image frames and weighting and fusing them with the statistical features of the current frame, the impact of instantaneous noise and brief motion events on motion intensity calculation is effectively smoothed. This temporal information integration makes the motion change intensity values ​​of each image sub-region more stable, accurate, and temporally continuous, thus providing a more reliable input for subsequent dynamic detection frequency planning. This helps ensure that autonomous vehicles can rationally allocate detection resources based on more accurate motion assessments in complex dynamic environments, improving the efficiency and accuracy of target detection.

[0048] In one feasible implementation, the step of using the vehicle speed data as input, querying a preset vehicle speed-frequency mapping table, and determining and outputting a global detection frequency reference value includes: determining the vehicle speed range to which the vehicle speed data belongs; querying the reference frequency value corresponding to the vehicle speed range from the preset vehicle speed-frequency mapping table, and using the reference frequency value as the output global detection frequency reference value; wherein, the preset vehicle speed-frequency mapping table contains multiple discrete vehicle speed ranges and their corresponding reference frequency values.

[0049] In this embodiment, when determining the global detection frequency benchmark value, it is first necessary to determine the vehicle speed range to which the current vehicle speed data belongs. This process aims to discretize continuous vehicle speed values ​​and classify them into several predefined speed ranges. For example, vehicle speed can be divided into multiple ranges such as "stationary," "low speed," "medium speed," and "high speed," each corresponding to a specific speed range, such as 0-10 km / h, 10-40 km / h, 40-80 km / h, and above 80 km / h. This determination can be achieved through a series of conditional statements (such as if-else if structures), comparing the current vehicle speed with these preset range boundaries to determine its affiliation. In this way, the system can simplify complex continuous speed information into more manageable discrete states, providing a clear index for subsequent frequency queries.

[0050] In this embodiment, the system then queries a preset vehicle speed-frequency mapping table for the reference frequency value corresponding to the current vehicle speed range. The vehicle speed-frequency mapping table is a pre-configured data structure whose core function is to establish a relationship between discrete vehicle speed ranges and their corresponding reference frequency values. This mapping table can exist in various forms, such as a lookup table stored in memory, a set of key-value pairs in a configuration file, or a record in a database. Each vehicle speed range uniquely corresponds to a reference frequency value, which represents the global default frequency used by the system for target detection within that speed range. For example, when the vehicle is in a low-speed range, the mapping table may specify a higher reference frequency to ensure a faster response to slow-moving targets such as pedestrians and bicycles; while when the vehicle is in a high-speed range, the mapping table may specify a relatively lower reference frequency that still meets safety requirements to balance computational resource consumption. This mapping relationship is usually set empirically or through optimization based on a large amount of actual driving data, simulation tests, and safety regulations. The query process involves finding the corresponding reference frequency value in the mapping table based on the determined vehicle speed range and outputting it as the global detection frequency reference value for subsequent local modulation processing.

[0051] In this embodiment, the above-described technical solution effectively addresses the issues of insufficient accuracy and lack of flexibility in determining the global detection frequency benchmark. By discretizing continuous vehicle speed data into defined speed ranges and querying a pre-defined speed-frequency mapping table, the system can dynamically and accurately obtain the most suitable global detection frequency benchmark for the current scenario based on the actual operating state of the vehicle. This refined mapping mechanism based on speed ranges ensures that the setting of the global detection frequency is no longer a simple linear relationship or a fixed value, but can be intelligently adjusted according to different driving scenarios. For example, in congested urban areas where vehicle speeds are low, the system can automatically query and adopt a higher benchmark frequency, thereby improving the real-time detection of surrounding dynamic targets and ensuring driving safety. In scenarios such as highways where vehicle speeds are high, the system can adopt a relatively lower benchmark frequency that still meets the detection requirements, effectively reducing unnecessary computational burden and optimizing system resource utilization. This method not only improves the efficiency and accuracy of target detection but also enhances the system's adaptability and robustness, providing a guarantee for the safe driving of autonomous vehicles in complex and ever-changing environments.

[0052] In one feasible implementation, the step of using the motion change intensity value and the global detection frequency reference value as inputs to perform local modulation processing and calculate and output the corresponding dynamic detection frequency value for each image sub-region includes: querying a preset motion intensity-modulation coefficient mapping relationship based on the motion change intensity value of each image sub-region to obtain and output the corresponding local modulation coefficient; multiplying the global detection frequency reference value by the local modulation coefficient to obtain and output the preliminary frequency value of the image sub-region; applying upper and lower limits to the preliminary frequency value and outputting the constrained value as the dynamic detection frequency value of the image sub-region; wherein, in the preset motion intensity-modulation coefficient mapping relationship, the motion change intensity value is proportional to the corresponding modulation coefficient.

[0053] In this embodiment, based on the motion intensity value of each image sub-region, a preset motion intensity-modulation coefficient mapping relationship is queried to obtain and output the corresponding local modulation coefficient. This aims to introduce an adjustment factor based on the local motion characteristics of each image sub-region. The motion intensity-modulation coefficient mapping relationship can be established in advance using experimental data, empirical rules, or machine learning methods, and stored as a lookup table, piecewise function, or continuous function. For example, when the motion intensity value is low, the corresponding local modulation coefficient may be close to 1, indicating a small impact on the global detection frequency benchmark; while when the motion intensity value is high, the corresponding local modulation coefficient will be significantly greater than 1, thereby increasing the detection frequency of that region. In the mapping relationship, the motion intensity value is directly proportional to the corresponding modulation coefficient, meaning that the more intense the motion in a region, the larger its local modulation coefficient, thus achieving a higher detection frequency in subsequent calculations. This proportional relationship can be linear or non-linear, as long as the modulation coefficient increases with increasing motion intensity.

[0054] In this embodiment, multiplying the global detection frequency reference value by the local modulation coefficients to obtain and output the preliminary frequency value of the image sub-region is a key step in realizing local frequency modulation. The global detection frequency reference value reflects the macroscopic requirements of the overall vehicle motion state for the detection frequency, while the local modulation coefficients reflect the dynamic activity level within the image sub-region. By multiplying the two, the preliminary frequency value of each image sub-region can take into account both the overall vehicle motion and the motion characteristics of the region itself, thereby achieving an effective combination of global and local factors.

[0055] In this embodiment, upper and lower bound constraints are applied to the initial frequency values, and the constrained values ​​are output as the dynamic detection frequency values ​​for the image sub-regions. This ensures that the calculated dynamic detection frequency values ​​are reasonable and feasible in practical applications. The upper bound constraint prevents excessively high detection frequencies, avoiding unnecessary waste of computational resources and preventing potential system instability or overload due to excessively high frequencies. For example, the upper bound can be set to the limit of the sensor frame rate or the system processing capacity. The lower bound constraint ensures that even in regions with very low motion intensity, a minimum detection frequency is maintained to avoid missing slowly moving or newly appearing objects, ensuring basic detection coverage and robustness. For example, the lower bound can be set to a frequency value sufficient for periodic checks. By applying these constraints, the dynamic detection frequency values ​​are ensured to be within an effective range, thereby optimizing system performance and resource utilization.

[0056] In this embodiment, the present application can perform refined and adaptive local modulation of the detection frequency based on the actual motion activity level of each image sub-region and the overall speed of the vehicle. This modulation mechanism avoids the unreasonable frequency allocation problems that may result from simply using a single global frequency or relying solely on local motion intensity. By introducing local modulation coefficients, the detection frequency can more accurately reflect the dynamic characteristics of different regions, ensuring higher detection priority for high-dynamic regions and appropriately lowering the frequency for low-dynamic regions. This effectively saves computational resources while maintaining detection performance. Furthermore, applying upper and lower bound constraints to the initial frequency values ​​further ensures that the allocated dynamic detection frequency values ​​are always within a reasonable and operable range, avoiding resource waste due to excessively high frequencies and the risk of missed detections due to excessively low frequencies, significantly improving the robustness and efficiency of the autonomous vehicle target detection system.

[0057] In one feasible implementation, the steps of taking the current image frame, the dynamic detection frequency planning data, and the historical image frames as inputs, inputting them to an adaptive sparse decision module for processing, and outputting a mask of the region to be processed with the same size as the current image frame include: parsing the dynamic detection frequency planning data, identifying all image sub-regions planned to be detected in the current frame, and marking the corresponding spatial positions in the mask of the region to be processed as a first identifier value; performing a spatiotemporal content consistency check on image sub-regions planned not to be detected in the current frame from the dynamic detection frequency planning data; and integrating the first identifier value and the updated identifier value after verification to form and output the complete mask of the region to be processed.

[0058] In this embodiment, the dynamic detection frequency planning data is first parsed to identify all image sub-regions planned to be detected in the current frame. For these regions, their corresponding spatial locations in the mask of the region to be processed are directly marked as the first identifier value, indicating that they are regions that clearly need to be detected. Subsequently, for those image sub-regions planned in the dynamic detection frequency planning data as not needing to be detected in the current frame, this application does not directly exclude them, but further performs spatiotemporal content consistency verification. This verification process is crucial; it determines whether these regions are truly in a stable state or whether there is dynamic information missed by the initial planning data by deeply analyzing the content changes of these regions in the time dimension. Specifically, the system extracts the current image block of the image sub-region that does not need to be detected from the current image frame, and extracts the historical image block sequence corresponding to the image sub-region from multiple historical image frames. Then, the content difference degree between the current image block and each image block in the historical image block sequence is calculated, and the comprehensive content difference degree value is aggregated and output. If the overall content difference value exceeds a preset consistency threshold, it indicates that the region has undergone significant temporal changes, potentially indicating the presence of new dynamic targets or important events. Therefore, its corresponding position in the mask of the region to be processed is updated and marked with the first identifier value, thus including it in the subsequent detection range. Conversely, if the overall content difference value does not exceed the consistency threshold, it indicates that the content of the region remains relatively stable over time and can be confirmed as not requiring detection in the current frame. Its corresponding position in the mask of the region to be processed is then marked with the second identifier value. Finally, this application integrates the initially marked first identifier value and the identifier value updated after spatiotemporal content consistency verification to form and output the complete mask of the region to be processed. This mask is a binary matrix with the same size as the current image frame, where the first identifier value represents the region to be detected, and the second identifier value represents the region not to be detected.

[0059] Through the above technical solution, this application introduces a spatiotemporal content consistency verification mechanism for initially "no-detection" areas based on dynamic detection frequency planning data. This enables the system to adaptively correct and optimize the initial detection priority division, effectively avoiding potential missed detections due to insufficient macroscopic motion intensity or lagging planning data. Even if the overall motion of a certain area is not obvious, as long as its content changes significantly over time, the system can promptly identify it and include it in the detection range. This greatly enhances the robustness and accuracy of target detection, especially when autonomous vehicles face complex, dynamic, and ever-changing traffic scenarios, enabling them to capture all key dynamic targets more comprehensively and reliably, thereby improving the environmental perception capability and decision-making safety of the autonomous driving system.

[0060] In one feasible implementation, for an image sub-region planned in the dynamic detection frequency planning data as not requiring detection in the current frame, the step of performing spatiotemporal content consistency verification includes: extracting the current image block of the image sub-region that does not require detection from the current image frame, and extracting the historical image block sequence corresponding to the image sub-region from multiple historical image frames; calculating the content difference degree between the current image block and each image block in the historical image block sequence, and aggregating and outputting a comprehensive content difference degree value; if the comprehensive content difference degree value exceeds a preset consistency threshold, then updating and marking the position of the image sub-region in the mask of the region to be processed with the first identifier value; if the comprehensive content difference degree value does not exceed the consistency threshold, then marking its position in the mask of the region to be processed with the second identifier value.

[0061] In this embodiment, the current image block of the image sub-region that does not require detection is first extracted from the current image frame, and a sequence of historical image blocks corresponding to the image sub-region is extracted from multiple historical image frames. This step aims to obtain the visual information of the region to be verified at the current moment and compare it with the visual information of the region over a past period. Specifically, the system accurately crops the corresponding image block from the current image frame based on the spatial coordinates of the "no-detection" region identified in the dynamic detection frequency planning data, and uses it as the current image block. Simultaneously, for temporal comparison, the system maintains a buffer of historical image frames, extracting image blocks from this buffer that correspond to the spatial position of the current image block from multiple consecutive historical image frames, forming a sequence of historical image blocks. These sequences of historical image blocks provide information on the evolution of the region over time, serving as the basis for determining whether its content has changed significantly.

[0062] In this embodiment, the content difference between the current image patch and each image patch in the historical image patch sequence is then calculated, and a comprehensive content difference value is aggregated and output. This step is used to quantify the degree of visual content change between the current image patch and the historical image patch sequence. The content difference can reflect whether there are newly appearing objects in the image patch, whether existing objects have moved significantly or disappeared, whether the lighting conditions have changed drastically, etc. The calculation method may include, but is not limited to: pixel-level difference (such as mean squared error, absolute difference sum), structural similarity index (SSIM), feature point matching difference (such as the distance after SIFT, SURF feature matching), or distance based on deep learning feature vectors (such as calculating cosine similarity or Euclidean distance after extracting features through a pre-trained convolutional neural network). Aggregating and outputting a comprehensive content difference value means comprehensively considering the difference between the current image patch and each historical image patch in the historical image patch sequence, for example, taking the average, maximum, weighted average, etc., to obtain a single value that can represent the overall spatiotemporal content change degree of the region.

[0063] In this embodiment, if the overall content difference value exceeds a preset consistency threshold, the position of the image sub-region corresponding to the mask of the region to be processed is updated and marked with the first identifier value. The consistency threshold is a preset value used to define whether the content difference has reached the level of "significant change". When the overall content difference value exceeds this threshold, it indicates that the visual content of the image sub-region has undergone a non-negligible change compared to the historical state. This may mean that a new dynamic target has appeared, or that the original scene has changed significantly. Even if the dynamic detection frequency planning data initially marked it as "not requiring detection", it should be reconsidered for detection. Therefore, the system updates the mask of the region to be processed, marking the position corresponding to the image sub-region with the first identifier value, indicating that it needs to be detected in the current frame.

[0064] In this embodiment, if the overall content difference value does not exceed the consistency threshold, its corresponding position in the mask of the region to be processed is marked as a second identifier value. When the overall content difference value does not exceed the preset consistency threshold, it indicates that the visual content of the image sub-region maintains a high degree of consistency between the current frame and historical frames, that is, no significant change has occurred. This verifies the initial judgment of the dynamic detection frequency planning data, that is, the region can indeed be regarded as low priority or does not need to be detected immediately in the current frame. Therefore, the system marks its corresponding position in the mask of the region to be processed as a second identifier value, indicating that it does not need to undergo detailed target detection processing in the current frame, thereby effectively saving computing resources.

[0065] In this embodiment, using the above technical solution, when certain image sub-regions are designated as not requiring detection by dynamic detection frequency planning data, this application does not simply ignore them, but introduces a spatiotemporal content consistency verification mechanism. This mechanism intelligently identifies regions that, although initially designated as not requiring detection, have undergone significant content changes by comparing the content differences between the current image block and historical image block sequences. For example, when a previously stationary vehicle suddenly starts moving, or a pedestrian suddenly enters a previously empty area, even if these regions were assigned a low detection priority in the dynamic detection frequency planning, this solution can promptly detect these abnormal changes. If the difference exceeds a preset threshold, it is forcibly marked as the first identifier value requiring detection in the mask of the region to be processed, ensuring that these sudden or important dynamic targets are not missed. Conversely, if the difference does not exceed the threshold, it is confirmed as the second identifier value, maintaining its state of not requiring detection, thereby avoiding redundant detection of truly static or low-dynamic regions. This adaptive verification strategy effectively compensates for the risk of missed detections that may be caused by relying solely on dynamic detection frequency planning data, significantly improves the robustness and accuracy of target detection in complex dynamic scenarios for autonomous vehicles, and avoids the huge computational burden caused by high-frequency detection of all areas, thus realizing the intelligent optimization and allocation of detection resources.

[0066] In one feasible implementation, the step of calculating the content difference between the current image patch and each image patch in the historical image patch sequence, and aggregating and outputting a comprehensive content difference value, includes: performing feature extraction on the current image patch to output a current feature vector; performing feature extraction on each historical image patch in the historical image patch sequence to output multiple historical feature vectors; calculating the distance between the current feature vector and each historical feature vector to obtain multiple corresponding distance values; averaging the multiple distance values ​​and outputting the average value as the comprehensive content difference value.

[0067] In this embodiment, when extracting features from the current image patch and outputting the current feature vector, the purpose is to obtain the essential visual information of the image patch, rather than directly comparing the original pixel values. Feature extraction provides a more robust and semantically meaningful representation of the image patch, effectively addressing the impact of factors such as illumination changes, noise interference, or slight deformation on image content comparison. Feature extraction methods can include, but are not limited to: using traditional feature descriptors such as SIFT (Scale Invariant Feature Transform), SURF (Speed-Up Robust Features), or ORB (Oriented Fast and Rotated BRIEF) to extract keypoints and their descriptors, and concatenating these descriptors or generating feature vectors through a bag-of-words model; or utilizing pre-trained convolutional neural networks (CNNs), such as VGG, ResNet, or MobileNet, to extract deep features from their intermediate layers and use these features as the current feature vector. These methods can capture the texture, edges, color distribution, or higher-level semantic information of the image patch.

[0068] In this embodiment, feature extraction is then performed on each historical image block in the historical image block sequence, resulting in multiple historical feature vectors. This step is consistent with the feature extraction method for the current image block, ensuring the comparability of the current image block and historical image blocks in the feature space. By independently extracting features from each historical image block, a series of feature vectors representing the content state of the region at different historical moments can be obtained, thereby constructing the historical trajectory of the evolution of the region's content over time.

[0069] Based on this, the distance between the current feature vector and each of the historical feature vectors is calculated, resulting in multiple corresponding distance values. The purpose of distance calculation is to quantify the degree of content difference between the current image patch and each historical image patch. Commonly used distance metrics include Euclidean distance, cosine distance (or 1 minus cosine similarity), Manhattan distance, etc. For example, Euclidean distance can intuitively represent the geometric distance of feature vectors in multidimensional space, while cosine distance focuses more on the similarity in vector direction. Choosing an appropriate distance metric can more accurately reflect the nature of image patch content changes.

[0070] In this embodiment, the multiple distance values ​​are finally averaged, and the average value is output as the comprehensive content difference value. By averaging the distance values ​​from different historical frames, the instantaneous noise or random changes that may exist in a single frame image can be effectively smoothed, thereby obtaining a more stable and representative comprehensive content difference value. This average value can more reliably reflect the overall content consistency or difference between the current image patch and the historical image patch sequence, providing a solid data foundation for subsequent spatiotemporal content consistency verification.

[0071] In this embodiment, through the above technical solution, this application can more accurately and robustly quantify the content differences between image blocks. The feature-based comparison method can effectively filter out interference from non-substantial factors such as image noise and illumination changes, thereby more accurately reflecting the actual changes in regional content. Furthermore, by averaging the distance values ​​of multiple historical frames, a more stable and reliable comprehensive content difference value can be obtained, effectively avoiding misjudgments caused by random factors in a single frame. This makes the spatiotemporal content consistency verification process more reliable, accurately identifying image sub-regions that, although initially planned as not requiring detection, have actually undergone significant content changes, and re-marking them as the first identifier value, i.e., the regions to be detected. Ultimately, this helps generate a more accurate mask for the regions to be processed, ensuring that the target detection neural network only processes the regions that truly need to be detected, thereby optimizing the allocation of computing resources and improving the efficiency and accuracy of the overall detection method.

[0072] In one feasible implementation, the method further includes: comparing the dynamic detection frequency planning data with the mask of the region to be processed, identifying all image sub-regions that are planned as not requiring detection in the dynamic detection frequency planning data but are marked as requiring detection in the mask of the region to be processed, and obtaining a set of difference regions; performing spatiotemporal distribution statistics on the set of difference regions, calculating their aggregation intensity in the image frame, and their occurrence frequency as the vehicle speed changes; determining whether the decision logic for generating the two-dimensional Boolean decision matrix needs to be optimized based on the aggregation intensity and the occurrence frequency; if it is determined that optimization is needed, generating a corresponding optimization instruction to adjust the decision logic.

[0073] In this embodiment, the dynamic detection frequency planning data is initially determined based on motion analysis and vehicle speed data, indicating which regions in the current frame should be detected. The region mask, after adaptive sparse decision processing such as spatiotemporal content consistency verification, ultimately determines the regions requiring target detection. By comparing the corresponding values ​​of these two matrices pixel-by-pixel or image sub-region-by-image, regions initially deemed unnecessary by the dynamic detection frequency planning data but subsequently marked as requiring detection by the region mask can be identified. These inconsistent regions are categorized into a set of discrepancies, directly reflecting the deviation between the system's initial decision logic and actual detection requirements. For example, pixels or sub-regions with a value of 0 (no detection required) in the dynamic detection frequency planning data and a value of 1 (detection required) in the region mask can be added to the set of discrepancies using logical XOR operations or simple conditional judgments.

[0074] In this embodiment, spatiotemporal distribution statistical analysis of the obtained set of difference regions aims to gain a deeper understanding of the patterns and regularities of these differences. On one hand, the clustering intensity of difference regions in the image frame is calculated. Clustering intensity reflects the degree of spatial concentration of difference regions. If difference regions exhibit obvious clustering in the image, rather than random dispersion, it indicates that the decision-making logic may have a systematic bias in a specific spatial region. For example, the clustering intensity can be quantified by calculating the number of connected components, the area of ​​each connected component, the centroid distribution, or by using clustering algorithms (such as DBSCAN, K-means). On the other hand, the frequency of occurrence of difference regions with varying vehicle speeds is calculated. This helps identify whether difference regions frequently appear within a specific speed range. If difference regions frequently appear within a specific speed range, it indicates that the decision-making logic may be insufficiently adaptable to the scene within that speed range. For example, the set of difference regions in historical frames can be associated with and stored with the corresponding vehicle speed data. Then, the number of occurrences or the total area of ​​difference regions in different speed ranges (e.g., 0-20km / h, 20-50km / h, and greater than 50km / h) can be counted to calculate their frequency in each speed range.

[0075] In this embodiment, based on the calculated clustering strength and frequency of occurrence, it is determined whether the decision logic for generating the two-dimensional Boolean decision matrix needs optimization. By comprehensively analyzing the statistical characteristics of the difference regions, the system can evaluate the reliability and adaptability of the current decision logic. If the differences are frequent and concentrated, or occur frequently under specific conditions (such as a specific vehicle speed), it indicates that the decision logic has defects and needs optimization. For example, a series of thresholds and rules can be set. If the clustering strength exceeds a preset threshold A (e.g., the percentage of the total image area occupied by the difference region, or the number of pixels in the largest connected region), and the frequency of occurrence exceeds a preset threshold B (e.g., the proportion of frames in which the difference region appears in the past N frames), it is determined that optimization is needed. Alternatively, a machine learning model can be used, taking the clustering strength, frequency of occurrence, and other relevant contextual information as input, to train the model to determine whether optimization is needed.

[0076] In this embodiment, if optimization is deemed necessary, a corresponding optimization instruction is generated to adjust the decision logic. When the system identifies a problem with the decision logic, action must be taken to correct it. The optimization instruction is a specific guide for the system on how to adjust the decision logic. These instructions can take various forms, such as adjusting thresholds used in dynamic region priority partitioning (e.g., frequency thresholds, weights of motion change intensity values, etc.), values ​​in the vehicle speed-frequency mapping table, or modifying or adding new rules to handle specific scenarios or vehicle speed ranges. If the decision logic is based on a machine learning model, retraining of the model can be triggered, using new data containing information about differing regions for learning. In some complex cases, the system may not be able to optimize automatically; in such cases, a prompt message can be generated to request manual intervention for analysis and adjustment.

[0077] In this embodiment, by introducing a real-time monitoring and analysis mechanism for the differences between dynamic detection frequency planning data and the mask of the region to be processed, this application can promptly identify and quantify potential deficiencies in the decision logic of dynamic region priority partitioning. Specifically, by performing spatiotemporal distribution statistics on the set of differing regions and calculating their aggregation intensity and frequency of occurrence as a function of vehicle speed, the system can intelligently determine whether the current decision logic needs optimization. When a systematic deviation in the decision logic is detected or its performance is poor under specific operating conditions, the system can automatically generate corresponding optimization instructions to adjust the decision logic. This enables the dynamic focus target detection method to possess adaptive learning and continuous optimization capabilities, effectively avoiding detection blind spots or redundant detection caused by fixed decision logic, thereby significantly improving the target detection efficiency, accuracy, and robustness of autonomous vehicles in complex dynamic environments.

[0078] In the embodiments of this application, the dynamic focus target detection method for autonomous vehicles acquires vehicle images and speed data, generates image motion vector fields based on motion analysis, performs dynamic region priority division in combination with vehicle speed, outputs dynamic detection frequency planning data, generates a mask for the region to be processed using an adaptive sparse decision module, selectively crops the image and inputs it into the target detection neural network, and finally outputs the results by mapping coordinates. This achieves intelligent allocation of computing resources, which can efficiently allocate computing resources, reduce processing latency, and improve the accuracy and real-time performance of key target detection.

[0079] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the dynamic focus target detection method for autonomous vehicles. Any simple modifications based on this technical concept are within the protection scope of this application.

[0080] This application also provides a dynamic focus target detection system for autonomous vehicles, with reference to... Figure 2The dynamic focus target detection system for autonomous vehicles includes: a memory 10, a processor 20, and a dynamic focus target detection program for autonomous vehicles stored on the memory 10 and executable on the processor 20. The dynamic focus target detection program for autonomous vehicles is configured to implement the steps of the dynamic focus target detection method for autonomous vehicles.

[0081] The dynamic focus target detection system for autonomous vehicles provided in this application employs the dynamic focus target detection method for autonomous vehicles described in the above embodiments, which can efficiently allocate computing resources, reduce processing latency, and improve the accuracy and real-time performance of key target detection. Compared with the prior art, the beneficial effects of the dynamic focus target detection system for autonomous vehicles provided in this application are the same as those of the dynamic focus target detection method for autonomous vehicles provided in the above embodiments, and other technical features of the dynamic focus target detection system for autonomous vehicles are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0082] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0083] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. All equivalent structural transformations made under the technical concept of this application using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of this application.

Claims

1. A dynamic focal target detection method for autonomous vehicles, characterized in that, The method includes: Acquire current image frames captured by the autonomous vehicle during its operation, as well as vehicle speed data from the vehicle bus interface; Motion analysis is performed based on the current image frame and at least one historical image frame to output an image motion vector field that reflects the direction and magnitude of pixel motion. Using the image motion vector field and the vehicle speed data as input, dynamic region priority partitioning is performed, and dynamic detection frequency planning data is output. The dynamic detection frequency planning data is used to define a two-dimensional Boolean decision matrix for whether different spatial regions in the current image frame are processed in this detection. The current image frame, the dynamic detection frequency planning data, and the historical image frames are used as inputs to the adaptive sparse decision module for processing, and the output is a mask of the unprocessed region with the same size as the current image frame; the mask of the unprocessed region is used as a binary matrix to identify the pixel region to be detected. Using the mask of the region to be processed and the current image frame as input, selective region cropping and stitching are performed, and the target candidate region image is output, which is composed of the image content of the region to be detected by the mask. The target candidate region image is input into the target detection neural network for processing, and an initial target detection result is output; the initial target detection result includes the category of at least one target and its bounding box coordinates in the target candidate region image; Using the initial target detection result and the spatial location mapping relationship recorded by the mask of the region to be processed as input, coordinate inverse mapping processing is performed to map the target bounding box coordinates to the original coordinate system of the current image frame, and the final target detection result is output. The step of using the image motion vector field and the vehicle speed data as input to perform dynamic region priority partitioning and output dynamic detection frequency planning data includes: The current image frame is divided into multiple image sub-regions, and the image motion vector field is used as input to calculate and output the motion change intensity value corresponding to each image sub-region. Using the vehicle speed data as input, a preset vehicle speed-frequency mapping table is queried to determine and output the global detection frequency reference value; Using the motion change intensity value and the global detection frequency reference value as input, local modulation processing is performed to calculate and output the corresponding dynamic detection frequency value for each image sub-region. The dynamic detection frequency value of each image sub-region is compared with a preset frequency threshold. If the dynamic detection frequency value of a sub-region is higher than or equal to the preset frequency threshold, it indicates that the sub-region needs to be processed for target detection in the current frame. Conversely, if it is lower than the preset frequency threshold, it indicates that the sub-region does not need to be detected in the current frame, or the detection priority is reduced, so as to generate and output the dynamic detection frequency planning data. The steps of using the motion change intensity value and the global detection frequency reference value as input, performing local modulation processing, and calculating and outputting the corresponding dynamic detection frequency value for each image sub-region include: Based on the motion change intensity value of each image sub-region, the preset motion intensity-modulation coefficient mapping relationship is queried to obtain and output the corresponding local modulation coefficient; The global detection frequency reference value is multiplied by the local modulation coefficient to obtain and output the preliminary frequency value of the image sub-region; Upper and lower limits are applied to the initial frequency value, and the constrained value is output as the dynamic detection frequency value of the image sub-region; wherein, in the preset motion intensity-modulation coefficient mapping relationship, the motion change intensity value is proportional to the corresponding modulation coefficient. The steps of taking the current image frame, the dynamic detection frequency planning data, and the historical image frames as inputs, feeding them into the adaptive sparse decision module for processing, and outputting a mask of the region to be processed with the same size as the current image frame include: The dynamic detection frequency planning data is analyzed to identify all image sub-regions that are planned to be detected in the current frame, and the corresponding spatial positions in the mask of the region to be processed are marked as the first identifier value. For the image sub-regions in the dynamic detection frequency planning data that are planned not to be detected in the current frame, perform spatiotemporal content consistency verification. Integrate the first identifier value and the updated identifier value after verification to form and output the complete mask of the region to be processed; For image sub-regions in the dynamic detection frequency planning data that are planned not to be detected in the current frame, the steps for performing spatiotemporal content consistency verification include: Extract the current image block of the image sub-region that does not need to be detected from the current image frame, and extract the historical image block sequence corresponding to the image sub-region from multiple historical image frames; Calculate the content difference between the current image block and each image block in the historical image block sequence, and aggregate and output a comprehensive content difference value; If the overall content difference value exceeds the preset consistency threshold, the position of the image sub-region in the mask of the region to be processed will be updated and marked with the first identifier value. If the overall content difference value does not exceed the consistency threshold, then its corresponding position in the mask of the area to be processed is marked as the second identifier value.

2. The dynamic focal target detection method for autonomous vehicles as described in claim 1, characterized in that, The steps of dividing the current image frame into multiple image sub-regions and calculating and outputting the motion change intensity value corresponding to each image sub-region, using the image motion vector field as input, include: For each image sub-region, extract all image motion vectors that fall within its boundary range; Calculate the amplitude of each extracted image motion vector to obtain the set of amplitude values ​​for that sub-region; Statistical analysis is performed on the set of amplitude values ​​to obtain the current frame statistical characteristics of the set; the statistical characteristics include at least one of the average amplitude, maximum amplitude, and standard deviation of amplitude. Read the historical motion change intensity value corresponding to the sub-region of the image from multiple consecutive historical image frames; The statistical features of the current frame are weighted and fused with the historical motion change intensity values, and the fusion result is output as the final motion change intensity value of the image sub-region.

3. The dynamic focus target detection method for autonomous vehicles as described in claim 1, characterized in that, The steps of using the vehicle speed data as input, querying a preset vehicle speed-frequency mapping table, determining and outputting the global detection frequency reference value include: Determine the vehicle speed range to which the vehicle speed data belongs; The reference frequency value corresponding to the vehicle speed range is retrieved from the preset vehicle speed-frequency mapping table, and the reference frequency value is used as the output global detection frequency reference value; wherein, the preset vehicle speed-frequency mapping table contains multiple discrete vehicle speed ranges and their corresponding reference frequency values.

4. The dynamic focal target detection method for autonomous vehicles as described in claim 1, characterized in that, The steps of calculating the content difference between the current image patch and each image patch in the historical image patch sequence, and aggregating and outputting a comprehensive content difference value, include: Perform feature extraction on the current image patch and output the current feature vector; Feature extraction is performed on each historical image block in the historical image block sequence, and multiple historical feature vectors are output accordingly. Calculate the distance between the current feature vector and each of the historical feature vectors respectively to obtain multiple corresponding distance values; The average value of the multiple distance values ​​is calculated and the average value is output as the comprehensive content difference value.

5. The dynamic focus target detection method for autonomous vehicles as described in claim 1, characterized in that, The method further includes: By comparing the dynamic detection frequency planning data with the mask of the region to be processed, all image sub-regions that are planned as not needing detection in the dynamic detection frequency planning data but are marked as needing detection in the mask of the region to be processed are identified, thus obtaining a set of difference regions. The spatiotemporal distribution statistics of the set of difference regions are performed to calculate their aggregation intensity in the image frame and their frequency of occurrence as a function of vehicle speed; wherein, the aggregation intensity is used to reflect the degree of spatial concentration of the difference regions. Based on the aggregation intensity and the occurrence frequency, determine whether the decision logic for generating the two-dimensional Boolean decision matrix needs optimization; if it is determined that optimization is needed, generate corresponding optimization instructions to adjust the decision logic.

6. A dynamic focus target detection system for autonomous vehicles, characterized in that, The dynamic focus target detection system for autonomous vehicles includes: a memory, a processor, and a dynamic focus target detection program for autonomous vehicles stored in the memory and executable on the processor, the dynamic focus target detection program for autonomous vehicles being configured to implement the steps of the dynamic focus target detection method for autonomous vehicles as described in any one of claims 1 to 5.

Images