A computer vision-based method and system for determining driving patterns

Abstract

This invention discloses a driving mode judgment method and system based on computer vision, belonging to the field of driving mode judgment technology. The method includes the following steps: extracting static image frames and dynamically adjusting the frame extraction frequency; performing local contrast adaptive adjustment on the image frames to obtain enhanced images; extracting and fusing brightness gradient and texture distribution based on multi-level visual feature analysis to generate a unified feature vector; inputting the unified feature vector into an intelligent discrimination model for navigation screen recognition and outputting results with confidence scores; and then dynamically weighting and updating the confidence scores of consecutive frames based on time series decay and a sliding window mechanism. This invention, through computer vision recognition and time series control, enables intelligent recognition and dynamic release of navigation screens by the vehicle terminal, avoiding interruptions caused by false restrictions, ensuring navigation continuity and driving safety, and maintaining recognition accuracy and response sensitivity in stable and abrupt scenarios through an adaptive adjustment mechanism, thereby improving the intelligence of the vehicle system and user experience.

Description

Technical Field

[0001] This invention relates to the field of driving mode determination technology, and specifically to a driving mode determination method and system based on computer vision. Background Technology

[0002] Using the phone mirroring function, drivers or passengers can project their mobile phone's application interface onto the vehicle's central control screen in real time, enabling seamless operation of various functions such as navigation, video playback, music control, communication, and voice assistants. However, to ensure driving safety, in-vehicle systems typically have safety restriction mechanisms. When video signals or moving images are detected in the mirrored screen, restriction policies are automatically triggered, such as pausing playback, obscuring the image, or issuing safety warnings. This mechanism aims to prevent drivers from being distracted by watching videos while driving, thereby reducing the risk of traffic accidents.

[0003] In the field of image recognition, navigation applications possess stable and unique visual features, such as road structure lines, directional arrows, path trajectories, map interactive elements, location information markers, and geographical names. These features exhibit a high degree of consistency at the image level. Leveraging computer vision techniques and algorithms, particularly open-source vision frameworks like OpenCV, preprocessing operations such as grayscale conversion, filtering, edge detection, and morphological analysis, combined with feature extraction algorithms like SIFT, ORB, or SURF, can be used to identify and classify mirrored images. Furthermore, machine learning models, such as SVM classifiers or lightweight convolutional neural networks (CNNs), can be used to train the extracted features, enabling automatic recognition of navigation images. This allows in-vehicle systems to distinguish between navigation applications and entertainment videos, providing intelligent support for driving safety control logic.

[0004] However, existing technologies have the following shortcomings:

[0005] Currently, most mainstream in-vehicle mirroring systems use video streaming for technical implementation. The in-vehicle terminal only receives continuous video frame data, not high-level data containing application semantic information. Therefore, the in-vehicle system cannot directly determine the source or type of the mirrored content; it can only make a rough identification based on the presence of a video signal. This method of judgment leads to the in-vehicle system's inability to distinguish between navigation screens and entertainment videos. While driving, when a user uses a mobile phone for navigation mirroring, the in-vehicle system may mistakenly identify the navigation screen as video playback, triggering safety restriction mechanisms and causing the screen to be obstructed or interrupted. This not only affects the continuous display of navigation information and the user experience but may also pose a potential hazard to driving safety.

[0006] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide a computer vision-based driving mode determination method and system to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a driving mode determination method based on computer vision, comprising the following steps:

[0009] Step 1: Based on the mirror video stream received by the vehicle system, a time segment sampling control mechanism is constructed to perform structured sampling on the mirror video stream, extract static image frames at fixed time intervals, and dynamically adjust the frame extraction frequency according to the resource load status of the vehicle system, so that the extracted static image frame sequence can ensure temporal representativeness while also having computational balance, and provide stable input data for subsequent image enhancement processing.

[0010] Step 2: Input the static image frames obtained from time segment sampling into the image enhancement processing flow, perform local contrast adaptive adjustment on the static image frames to enhance the geometric features of navigation elements, highlight road information such as road structure, directional arrows and path trajectory, and obtain an enhanced image with balanced visual features, providing a unified visual benchmark for feature extraction.

[0011] Step 3 enhances the image input feature extraction process by performing local structure recognition and global texture parsing based on multi-level visual feature analysis. This extracts multi-dimensional visual feature information, including road lines, directional arrows, path contours, and map symbols, and integrates brightness gradient and texture distribution information to generate a unified feature vector for content discrimination.

[0012] Step 4: Input the unified feature vector into the intelligent discrimination model, perform category recognition on the navigation screen through nonlinear feature mapping relationship, output the recognition result containing confidence weight, and adjust the recognition result with weight based on dynamic confidence correction strategy to form a high-confidence navigation screen discrimination output with time weight;

[0013] Step 5: Based on the recognition results with time weights output by the intelligent discrimination model, a time series decay and sliding window statistical mechanism is constructed to dynamically update the recognition status of multiple consecutive frames with weighted confidence. When the recognition results of consecutive frames are consistent, the system response update interval is extended, and when a sudden change in screen features is detected, the feedback cycle is shortened, thereby realizing adaptive release of navigation screens and closed-loop control of safety restrictions on non-navigation screens.

[0014] Preferably, the specific steps for constructing a time-segment sampling control mechanism based on the mirrored video stream received by the vehicle system are as follows:

[0015] After the vehicle terminal receives the mirrored video stream from the mobile device, it segments the mirrored video stream according to the timeline. By reading the timestamp information of the video frames, the video stream is divided into multiple consecutive time segments, and the video frames in each time segment are numbered and registered to generate a frame index table.

[0016] After the time segment is divided, static image frames are extracted from each time segment at fixed time intervals. The frame extraction frequency is dynamically adjusted according to the resource load status of the vehicle terminal. The frame extraction time interval is extended under high load and shortened under low load to maintain temporal representativeness and computational balance.

[0017] After dynamic frame extraction is completed, the extracted static image frames are organized and cached sequentially. The logical positions between frames are corrected through timestamp alignment, and a cache queue is established according to the first-in-first-out principle to ensure that the frame sequence in the cache is consistent in terms of time interval, providing continuous and stable input data for subsequent image enhancement processing.

[0018] Preferably, the static image frames obtained by sampling time segments are input into the image enhancement processing flow, and the specific steps for performing local contrast adaptive adjustment on the static image frames are as follows:

[0019] The static image frames obtained by sampling time segments are input into the first stage of the image enhancement process. Basic standardization processing is performed on each static image frame. Through size adjustment, brightness equalization and color temperature correction, all input image frames achieve a unified standard in terms of brightness, color temperature, resolution and aspect ratio.

[0020] After completing the basic standardization process, local region segmentation and partition brightness analysis are performed on the adjusted static image frames. Based on the brightness distribution range and contrast change characteristics of each region, differentiated brightness adjustment is performed. The brightness difference is enhanced in the region containing road elements, and the brightness of non-critical regions is adjusted smoothly to highlight the road structure and directional arrows.

[0021] After completing the local brightness partition adjustment, geometric enhancement and edge smoothing optimization are performed on the adjusted static image frames. By identifying the brightness gradient direction, the road main line and path trajectory boundary are enhanced, and the overlapping area is smoothed to obtain an enhanced image frame with a clear structure.

[0022] After completing the geometric enhancement, global visual equalization and continuous frame brightness smoothing are performed. The enhanced image frames are then redistributed and their continuity is corrected in both spatial and temporal dimensions to generate a continuous enhanced image sequence with consistent brightness and balanced contrast.

[0023] Preferably, in the process of global visual equalization and continuous frame brightness smoothing, the average brightness curves of the continuous enhanced image frames are dynamically aligned, and the brightness transition between adjacent frames is adjusted based on the time sequence of brightness distribution of each frame, so that the continuous enhanced image sequence remains smooth and consistent in the time dimension, thereby ensuring that the navigation elements have stable visual continuity and structural consistency in the inter-frame display process.

[0024] Preferably, the enhanced image input feature extraction process generates a unified feature vector for content discrimination through the following specific steps:

[0025] The enhanced image input feature extraction process divides the spatial distribution of the enhanced image and identifies key regions. By analyzing the brightness distribution, shape boundaries, and pixel arrangement direction, regions with structural regularities are identified, and their spatial location, shape contours, and connection relationships are recorded.

[0026] After obtaining the spatial distribution division and key area identification results, geometric structure analysis is performed on each local area to extract the boundary line features of the road area, the angle features of the directional arrows, the morphological features of the path trajectory, and the geometric features of the map symbols. The extracted geometric features are then spatially indexed and arranged to form local structural features.

[0027] After extracting local structural features, global visual feature analysis is performed on the enhanced image. By statistically analyzing the brightness gradient direction, color distribution, and texture density, global visual features reflecting the overall layout features of the navigation screen are formed, and the global visual features are aligned with the local structural features in spatial coordinates.

[0028] After obtaining local structural features and global visual features, the two are fused together. Different types of feature information are integrated into a unified feature vector according to spatial location and content weight, which is used to fully express the spatial geometric structure, brightness distribution pattern, texture arrangement features and element layout characteristics of the enhanced image.

[0029] Preferably, when fusing local structural features and global visual features, the road line direction information is matched with the global brightness distribution direction according to the spatial position correspondence, the path contour morphology parameters are associated with the global texture density, and the direction arrow angle information is combined with the brightness concentration trend to ensure that the unified feature vector generated by fusion maintains consistency in spatial geometric structure and visual distribution features.

[0030] Preferably, the unified feature vector is input into the intelligent discrimination model, and the navigation screen is classified through nonlinear feature mapping relationship. The specific steps are as follows:

[0031] The unified feature vector generated in the previous stage is input into the intelligent discrimination model. The input unified feature vector is initially classified and identified. The correspondence between features and navigation categories is established through feature normalization and multi-layer nonlinear mapping to obtain preliminary classification results.

[0032] After the initial identification by the intelligent discrimination model is completed, the confidence level and weight allocation of the initial classification results output by the intelligent discrimination model are performed. The confidence weight of each category is determined according to the discrimination strength and feature matching degree within the intelligent discrimination model, so that the identification results have a credible expression.

[0033] After obtaining the recognition result containing confidence weights, the recognition result is dynamically confidence corrected. The confidence weights are smoothly adjusted between adjacent image frames with time as a reference, and the persistence and response sensitivity of the recognition result are controlled by the time decay factor.

[0034] After completing the dynamic confidence correction, the corrected recognition results are subjected to time-weighted fusion processing. The recognition results with time weights are generated by weighted averaging within the time window, so that the output results are continuous in time and controllable in stability.

[0035] Preferably, the time-weighted fusion processing includes: dynamically adjusting the time weight allocation based on the confidence fluctuation of the recognition results of consecutive frames within the time window; increasing the weight of historical frames when the recognition results remain stable; and increasing the weight of the current frame when the recognition results change, so that the output comprehensive judgment result reflects the continuity of the screen content in the time dimension and has real-time response capability.

[0036] Preferably, based on the identification results with time weights output by the intelligent discrimination model, the specific steps for constructing the time series decay and sliding window statistical mechanism are as follows:

[0037] The recognition results with time weights output by the intelligent discrimination model are input into the time series management process. The recognition results are arranged sequentially and time-calibrated on the time axis. The timestamp, category label and confidence weight of each frame recognition result are recorded synchronously, and the continuity of the time series is maintained through time alignment processing.

[0038] After the time series is constructed, the sliding window statistical mechanism is used to dynamically analyze the recognition results of consecutive frames. A sliding window of a preset fixed length is set on the time axis, and the confidence weight and time position of the recognition results of each frame in the window are statistically judged to determine the stability or abruptness of the image state.

[0039] After obtaining the sliding window statistics, the recognition status within the window is dynamically updated with weighted confidence. The confidence of each frame is weighted according to the time sequence, and a confidence decay factor is introduced based on the confidence change rate to balance the temporal stability and response sensitivity of the recognition results.

[0040] After completing the dynamic confidence weighted update, the system response update interval and feedback cycle are adaptively adjusted according to the time series change trend. When the recognition results of consecutive frames are consistent, the system response update interval is extended, and when a sudden change in screen features is detected, the feedback cycle is shortened, thereby realizing adaptive release of navigation screen and closed-loop control of safety restriction of non-navigation screen.

[0041] The computer vision-based driving mode judgment system includes a time segment sampling control module, an image enhancement processing module, a feature extraction and analysis module, an intelligent discrimination and recognition module, and a time series dynamic update module.

[0042] The time segment sampling control module constructs a time segment sampling control mechanism based on the mirrored video stream received by the vehicle system, extracts static image frames at fixed time intervals, and dynamically adjusts the frame extraction frequency according to the resource load status of the vehicle system.

[0043] The image enhancement processing module inputs the static image frames obtained by sampling time segments into the image enhancement processing flow, performs local contrast adaptive adjustment on the static image frames, enhances the geometric features of navigation elements, and obtains an enhanced image with balanced visual features.

[0044] The feature extraction and analysis module will enhance the image input feature extraction process. Based on multi-level visual feature analysis, it will perform local structure recognition and global texture parsing, extract multi-dimensional visual feature information, and fuse brightness gradient and texture distribution information to generate a unified feature vector.

[0045] The intelligent discrimination and recognition module inputs a unified feature vector into the intelligent discrimination model, performs category recognition on the navigation screen through a nonlinear feature mapping relationship, outputs a recognition result containing confidence weights, and performs weighted adjustment on the recognition result based on a dynamic confidence correction strategy;

[0046] The time series dynamic update module constructs a time series decay and sliding window statistical mechanism based on the recognition results with time weights output by the intelligent discrimination model, and performs dynamic confidence weighted updates on the recognition status of multiple consecutive frames.

[0047] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0048] This invention introduces a computer vision-based content recognition mechanism into the in-vehicle terminal, enabling the system to automatically extract and analyze the visual features of the navigation screen from the mirrored video stream, achieving accurate recognition and dynamic release control of navigation-related screens. Through the coordinated processing of time-segment sampling and image enhancement, this solution allows the in-vehicle terminal to obtain high-quality, stable image input data even under limited computing power. This effectively avoids the problem of navigation screens being misinterpreted as video playback while driving, causing obstruction and interruption, thus ensuring continuous display of navigation information and driving safety.

[0049] This invention combines an intelligent discrimination model with a time-series sliding mechanism, enabling the recognition results to adaptively adjust in both spatial and temporal dimensions. The system can dynamically adjust its response frequency based on changes in the screen content, maintaining smooth output when recognition is stable and responding quickly to sudden changes in the screen. This achieves accurate differentiation and closed-loop control of navigation and non-navigation content, not only enhancing the intelligence and environmental adaptability of the in-vehicle system but also achieving an effective balance between driving safety and user experience. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0051] Figure 1 This is a flowchart of the method for determining driving modes based on computer vision according to the present invention;

[0052] Figure 2 This is a flowchart illustrating the time-segment sampling control mechanism for constructing a video stream received by an in-vehicle system according to the present invention.

[0053] Figure 3 This invention provides a flowchart for generating a unified feature vector from an enhanced image input feature extraction process.

[0054] Figure 4 This is a schematic diagram of the computer vision-based driving mode determination system of the present invention. Detailed Implementation

[0055] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0056] like Figures 1 to 3As shown, this invention provides a driving mode determination method based on computer vision, including the following steps:

[0057] Step 1: Based on the mirror video stream received by the vehicle system, a time segment sampling control mechanism is constructed to perform structured sampling on the mirror video stream, extract static image frames at fixed time intervals, and dynamically adjust the frame extraction frequency according to the resource load status of the vehicle system, so that the extracted static image frame sequence can ensure temporal representativeness while also having computational balance, and provide stable input data for subsequent image enhancement processing.

[0058] The mirrored video stream is structured and sampled to extract still image frames at fixed time intervals. The specific steps are as follows:

[0059] After receiving the mirrored video stream from the mobile device, the in-vehicle terminal performs precise segmentation processing according to the timeline. Specifically, during the continuous transmission of the mirrored video, the video stream is divided into multiple consecutive time segments by reading the timestamp information of the video frames. Each time segment represents a fixed-length time interval, such as 200 milliseconds to 500 milliseconds. A strict timestamp sorting method is used during segmentation to ensure that the order of each frame on the timeline is completely consistent with the actual playback order, avoiding misalignment caused by frame disorder or missing frames. The start and end boundaries of each time segment are determined by precise time control logic, ensuring seamless connection between consecutive time segments without overlapping or gaps in time intervals. After the time segmentation is completed, the video frames in each time segment are numbered and registered, generating a frame index table for precise location in subsequent frame extraction operations. Through this timestamp-based structured management method, the entire mirrored video stream is transformed into an ordered, hierarchical set of frame data in time, providing a clear logical basis for subsequent frame extraction and load adjustment.

[0060] After dividing the system into time segments, static image frames are extracted from each segment at fixed time intervals. The frame extraction frequency is dynamically adjusted based on the resource load status of the vehicle terminal. Specifically, when the system is running stably and computing resource utilization is within a normal range, image frames are extracted from consecutive time segments at a preset base time interval. For example, when the base time interval is set to 300 milliseconds, the system extracts one image frame at the middle of each time segment, ensuring that the extracted image frames are evenly distributed across the time dimension, thus providing good temporal representativeness. When the system's computing resources are detected to be under high load, such as when simultaneously performing operations like speech recognition, navigation path rendering, or human-machine interface refresh, the system automatically extends the frame extraction time interval to 500 milliseconds or higher based on real-time monitoring of CPU and memory utilization. This reduces the frame extraction frequency, decreases the performance overhead of image processing tasks, and prevents task delays due to insufficient computing resources. When the system is under low load, such as when the user is not multitasking or the vehicle is stationary, the system automatically shortens the frame extraction interval to 200 milliseconds or even less, thereby increasing the frame sampling density and making the recognition process more detailed in terms of temporal continuity. The dynamic adjustment is based not only on CPU and memory utilization, but also on the current latency level of the video frame buffer and the input rate of the video stream. The system comprehensively judges the current processing capacity margin based on these parameters and adjusts the frame extraction frequency in real time. Through this dynamic frame extraction strategy based on system load feedback, it is possible to effectively balance the allocation of computing resources while ensuring the temporal integrity of the video frame, making the frame extraction process both stable and real-time.

[0061] After dynamic frame extraction, the extracted static image frames are sequentially organized and cached to ensure a continuous flow of structurally stable input data in subsequent image enhancement stages. Specifically, the system numbers the extracted static image frames according to time segments, establishing a frame sequence index table and storing each frame image in a one-to-one correspondence with its corresponding timestamp. Frame data is cached in a fixed-length buffer queue according to a first-in, first-out (FIFO) principle. This buffer queue typically contains several consecutive time segments to maintain the temporal continuity required for recognition. When a new image frame is extracted, the system automatically removes the oldest frame data from the buffer queue, ensuring that the buffer maintains a fixed number of consecutive frames at any given time. To prevent uneven frame intervals due to network fluctuations or decoding delays during video stream transmission, a timestamp alignment operation is performed during the frame sequence organization stage. This involves correcting the logical position of frames in the buffer based on the actual time difference between frames, ensuring that the frame time interval in the buffer is strictly consistent with the preset sampling interval. This operation ensures that the interval between each frame on the logical time axis remains constant, thereby avoiding image distortion caused by sudden input changes or time jumps in subsequent image enhancement stages. Furthermore, to ensure continuous data input, the system immediately performs a data validity check after each frame extraction to determine whether the frame has been completely received and correctly stored. Only frames with complete data structure and consecutive timestamps are added to the buffer queue. Through this sequential and time-aligned management approach, subsequent image enhancement processes can continuously receive input image sequences with temporal coherence and spatial consistency.

[0062] Through the above steps, the entire time-segment sampling control process achieves precise segmentation of the mirrored video stream in the time dimension, real-time balanced control of the frame-sampling frequency in the computational dimension, and temporal organization and stable input of image frames in the data dimension. In this way, continuous mirrored images are transformed into a uniform, stable, and controllable static image sequence, preserving the dynamic temporal characteristics of the mirrored images while significantly reducing the computational burden required for real-time recognition, thus providing a reliable input data foundation for subsequent image enhancement and feature extraction. The entire process achieves coordinated control of the mirrored video stream in terms of temporal distribution, frame-sampling frequency, and input stability, enabling the driving mode judgment method to maintain efficient, stable, and continuous visual input capabilities even in complex in-vehicle operating environments.

[0063] Step 2: Input the static image frames obtained from time segment sampling into the image enhancement processing flow, perform local contrast adaptive adjustment on the static image frames to enhance the geometric features of navigation elements, highlight road information such as road structure, directional arrows and path trajectory, and obtain an enhanced image with balanced visual features, providing a unified visual benchmark for feature extraction.

[0064] The static image frames obtained from time segment sampling are input into the image enhancement processing flow. Local adaptive contrast adjustment is performed on the static image frames. The specific steps are as follows:

[0065] The first stage of the image enhancement process involves inputting static image frames obtained from time-segment sampling into the image enhancement workflow. These frames undergo a standardized process to ensure uniformity. Specifically, the image dimensions are first adjusted to maintain consistency in aspect ratio, pixel density, and boundary range across all frames. This ensures subsequent operations can be performed at the same spatial scale, preventing feature position shifts or proportional distortions caused by differences in output resolution across different devices. After size standardization, the brightness distribution of each frame is globally statistically analyzed to calculate the average brightness value. This value is then used as a reference to gradually increase the brightness of overly dark areas and moderately compress the brightness of overly bright areas, achieving a balanced distribution of overall light intensity. This prevents overexposure or excessively deep shadows caused by glare from car interiors or automatic brightness adjustments from mobile phones. Furthermore, the color channels are balanced based on the overall color temperature characteristics of the image, maintaining neutral warm and cool tones to prevent blue light enhancement during nighttime driving or yellow cast under strong daylight from affecting structural feature recognition. Through these processes, all input image frames achieve a unified standard in terms of brightness, color temperature, resolution, and aspect ratio, providing a stable foundation for subsequent, more refined local enhancement.

[0066] After basic standardization, local region segmentation and zonal brightness analysis are performed on the adjusted static image frames. Specifically, each frame is divided into multiple overlapping rectangular regions based on spatial distribution characteristics. The size of each region is determined according to the screen resolution and display ratio to ensure finer adjustments to detailed areas in high-resolution scenes. For each local region, the brightness distribution range and contrast variation are statistically analyzed, and the distribution and structural direction of bright and dark edges within the region are examined. In navigation screens, road lines are typically continuous, bright, thin lines, directional arrows have fixed shapes, and map paths often extend diagonally or along curves; therefore, the brightness variation direction in these areas is regular. Based on this characteristic, the brightness difference adjustment range is increased for areas containing road elements to make the brightness contrast between lines and background more obvious, thereby enhancing their visual recognizability. For map backgrounds, place name signs, or non-critical areas, a more gradual brightness adjustment range is used to avoid damaging the original color levels. The processing of each local region overlaps spatially to ensure continuous and smooth brightness changes at the region edges, avoiding bright spots or dark bands caused by excessive local enhancement. Through this process, the entire image forms a differentiated contrast distribution in a local area, which further highlights the road edges, arrow boundaries, and path trajectories visually, while the background remains soft, creating a distinct and natural enhancement effect.

[0067] After adjusting the local brightness zones, geometric enhancement and edge smoothing optimization are performed on the entire image. This step uses the enhanced image with local brightness adjustments as input, focusing on contour enhancement of navigation-related geometric structures. By identifying the direction of brightness gradient distribution in the image, the brightness difference between the edges and the background is gradually increased for linear and curved structural areas such as main road lines, turning boundaries, intersections, and directional arrows, making these structures more prominent in the image. For areas where road boundaries and path trajectories intersect in the navigation interface, the brightness distinction between the two is increased to ensure clear and unblurred intersection boundaries. Simultaneously, for scenes with complex layer overlays, such as road name text, navigation prompt windows, and icon coverage areas, a smooth transition is applied to edge areas to prevent local visual breaks during the enhancement process. This results in enhanced image frames where brightness changes gradually at the boundaries of areas, maintaining structural clarity while avoiding the appearance of broken lines caused by over-sharpening. After this step is completed, all navigation-related geometric elements in the image are enhanced, the path trajectory appears as a continuous bright line, the directional arrow boundaries are clear, the map structure is clearly layered, and a geometric visual result suitable for feature extraction is formed.

[0068] After geometric enhancement, global visual equalization and continuous frame brightness smoothing are performed on the entire frame image to ensure consistency in both spatial and temporal dimensions. Specifically, the overall brightness distribution of each enhanced image frame is analyzed, and the average brightness curve of all frames is used as a global reference. Adjustments are made to areas with excessively high or low brightness in individual frames to ensure a smooth transition in brightness between adjacent frames, avoiding visual discomfort such as flickering or abrupt brightness changes during continuous playback. Simultaneously, in the spatial dimension, the brightness histogram distribution of the entire image is smoothly redistributed, making the transition between local bright and dark areas more natural and ensuring consistent visual performance of the enhanced navigation elements across different frames. Furthermore, the globally equalized frames are rearranged according to their sampling time order to generate a continuous enhanced image sequence. This ensures that the enhancement results achieve structural equalization not only within a single frame but also maintain brightness consistency and visual continuity between frames. Through this dual spatial and temporal equalization approach, a unified visual standard is achieved for the entire enhanced image sequence in terms of brightness, contrast, and detail.

[0069] Through the above steps, the static image frames obtained from time-segment sampling, after basic standardization, local contrast adjustment, geometric structure enhancement, and global brightness equalization, form a set of enhanced images with clear visual features, balanced brightness, and stable structure. This enhanced image sequence maintains consistent contrast and detail in the representation of road structures, directional arrows, path trajectories, and map elements. It effectively copes with different lighting environments, display device differences, and dynamic scene changes, providing a high-quality visual input foundation for subsequent feature extraction and intelligent recognition. This ensures the stability and real-time performance of the driving mode judgment process under complex in-vehicle operating conditions.

[0070] Step 3 enhances the image input feature extraction process by performing local structure recognition and global texture parsing based on multi-level visual feature analysis. This extracts multi-dimensional visual feature information, including road lines, directional arrows, path contours, and map symbols, and integrates brightness gradient and texture distribution information to generate a unified feature vector for content discrimination.

[0071] The image input feature extraction process will be enhanced to generate a unified feature vector for content discrimination. The specific steps are as follows:

[0072] The enhanced image frame, processed through image enhancement (i.e., the enhanced image input feature extraction process), is used to divide the spatial distribution of the enhanced image and identify key regions. Specifically, based on the pixel resolution, display ratio, and content layout of the enhanced image, the entire image is divided into multiple equidistant and partially overlapping rectangular regions. The number of pixels covered by each region is dynamically determined according to the screen resolution to ensure that the region division captures local details while preserving the overall structural continuity. After division, the brightness distribution, shape boundaries, and pixel arrangement direction within each region are analyzed to identify regions with structural regularities. Navigation screens typically contain directional geometric features, such as road lines existing as continuous bright bands, directional arrows arranged at specific angles, path trajectories traversing the map area in arc or broken line form, and map symbols and place name markers concentrated in specific areas. Based on these regular visual features, regions with obvious directionality, clear boundaries, or repetitive structures are highlighted and their spatial location, shape outline, and connections to adjacent regions are recorded, ensuring that the spatial structural information of the image is completely preserved in a regionalized form. This partitioning recognition method structurally decomposes the enhanced image into several key regions with spatial dependencies, providing a clear localization basis for subsequent local structural feature analysis.

[0073] After obtaining the spatial distribution division and key area identification results, geometric structure analysis is performed on each local area to extract basic morphological features from the navigation screen. Specifically, in the road area, the brightness variation trend along the road extension direction is analyzed, continuous light and dark boundary lines are captured, and the width, length, and direction information of these lines are recorded as structural features. In the area containing directional arrows, the geometric relationship between the head and tail of the arrow is extracted, and its pointing angle and shape ratio are determined by identifying the brightness aggregation shape and edge sharpness difference of the arrow tip. In the path trajectory area, the path direction composed of continuous bright lines is identified, and a path morphological description is formed by detecting its curvature changes and direction. In the map symbol area, the outline features and arrangement rules of the symbols are identified, and their relative spacing, boundary shape, and repetition features are extracted. To ensure that the feature information of different areas remains consistent at the spatial level, all extracted geometric features are rearranged according to the area index to ensure the continuity of structural features between adjacent areas, thus forming a complete set of local structural features. This set includes the direction information of road lines, the morphological information of path outlines, the geometric features of arrow symbols, and the distribution features of map elements, providing accurate geometric references for subsequent extraction of global visual features.

[0074] After extracting local structural features, a global visual feature analysis is performed on the entire enhanced image to extract macroscopic features that reflect the overall style and content distribution of the navigation interface. Specifically, a global brightness distribution map is generated by summarizing and statistically analyzing the brightness gradient direction, color distribution, and texture density of all regions in the entire enhanced image, which describes the overall trend of brightness changes in the image. In navigation images, roads and paths typically exhibit continuous brightness bands, while the map background area displays relatively smooth texture features. Based on this difference, the concentration of brightness change directions and the consistency of texture arrangement in the image are calculated to identify regions with navigation features. For example, road lines and directional arrows usually show a consistent directional distribution in the brightness gradient direction, while place names and icons exhibit strong texture density. After integrating these global visual features, a global visual feature description reflecting the overall layout of the navigation image can be obtained, including indicators such as brightness continuity, directional consistency, texture density distribution, and color gradient balance. In addition, in order to maintain the correspondence with local structural features, the global visual features are aligned with the structural features of each region on the spatial coordinates, so that the overall feature description and local morphological features form a spatial mapping relationship, thereby realizing the basis for the fusion of local details and overall visual features.

[0075] After obtaining local structural features and global visual features, the two are fused to generate a unified feature vector for content discrimination. Specifically, the directional information of road lines is combined with the global brightness distribution direction to form a vector reflecting the consistency of navigation path direction; the morphological parameters of the path contour are fused with the global texture density to form morphological features describing the distribution of map areas; the angular information of directional arrows is correlated with the global brightness concentration trend to generate feature data representing the navigation direction; and the distribution features of map symbols are combined with the global color distribution to form a combination feature describing the arrangement rules of interface elements. During the fusion process, the proportional consistency and logical relationship between different features are maintained according to the spatial position and content weight of each feature, so that low-level local structural features and high-level global visual features jointly express the screen content in the same coordinate system. All fused feature information is integrated into a unified multi-dimensional vector, which fully reflects the spatial geometric structure, brightness distribution rules, texture arrangement features, and element layout characteristics of the enhanced image. Through this vector, the road structure, directional arrows, path trajectories, and map symbols in the navigation screen are expressed in an orderly manner in a unified feature space.

[0076] Through the sequential implementation of the above steps, the enhanced image achieves local detail extraction at the structural level, global visual feature analysis at the visual level, and multi-dimensional feature integration at the fusion level. The resulting unified feature vector accurately describes the spatial layout and visual attributes of the navigation screen. This feature vector exhibits stable feature representation capabilities under different navigation applications, lighting environments, and display styles, providing a clear, continuous, and identifiable visual foundation for driving mode judgment methods. This enables in-vehicle terminals to perform intelligent judgment and safety control based on image content in real-time environments.

[0077] Step 4: Input the unified feature vector into the intelligent discrimination model, perform category recognition on the navigation screen through nonlinear feature mapping relationship, output the recognition result containing confidence weight, and adjust the recognition result with weight based on dynamic confidence correction strategy to form a high-confidence navigation screen discrimination output with time weight;

[0078] A unified feature vector is input into the intelligent discrimination model, and the navigation screen is classified through a non-linear feature mapping relationship. The specific steps are as follows:

[0079] The unified feature vector generated in the previous stage is input into the intelligent discrimination model for initial classification and recognition of the input multidimensional feature information. Specifically, the unified feature vector includes geometric morphology features, brightness gradient features, texture distribution features, and spatial layout features extracted from the enhanced image. These features are used as input parameters into the intelligent discrimination model. The structure of the intelligent discrimination model includes multiple feature mapping layers, used to perform nonlinear transformations and high-dimensional spatial mapping on the input features to establish the correspondence between features and navigation categories. In this process, the model first normalizes each feature according to the dimension and type of the input vector to maintain consistency in numerical scale and avoid judgment bias caused by differences in feature dimensions. Subsequently, the model maps the input vector layer by layer according to the spatial distribution characteristics of the features to capture the inherent relationships between different visual elements in the navigation screen. For example, the correlation between the directional distribution of road lines and the changes in brightness gradient, and the correlation between the curvature of the path trajectory and the density of map symbols are all comprehensively considered in the feature mapping. Through this feature-based spatial mapping method, the intelligent discrimination model can establish nonlinear mapping relationships between complex features, thereby achieving preliminary classification and recognition of the input image content, and providing basic recognition results for subsequent confidence calculation and dynamic adjustment.

[0080] After the initial identification by the intelligent discrimination model, the output results are evaluated for confidence and weights to form a classification output including confidence weights. Specifically, after performing multi-layer feature mapping on the input unified feature vector, the intelligent discrimination model outputs recognition results corresponding to each category, including judgment values ​​for navigation and non-navigation categories. For this output, the confidence weight for each category is calculated based on the model's internal discrimination strength and feature matching degree. The confidence weight reflects the model's trustworthiness of the current recognition result, and its numerical range is determined by the model's internal response features. In this process, if the geometric structure contained in the input unified feature vector has high similarity to the navigation screen features—for example, continuous road boundaries, clear directional arrows, regular path trajectories, and balanced distribution of map symbols—then the confidence weight of the navigation category output by the model will be high. Conversely, if the screen lacks these structural features or contains highly dynamic content, the confidence weight of the model output will be correspondingly lower. This step ensures that the model's recognition results not only output category judgments but also include quantitative confidence parameters, providing a measurable reference for subsequent dynamic corrections. In this way, the intelligent discrimination model completes feature category recognition at the spatial level and establishes confidence expression at the numerical level, making the recognition results discriminative and reliable.

[0081] The following examples provide detailed illustrations. It should be understood that the numerical values ​​used in the examples are hypothetical examples for ease of understanding and do not represent the specific data in actual testing. This application will not elaborate on these figures further.

[0082] Assuming the vehicle is in motion, the driver projects the navigation screen onto the central control display using their phone's mirroring function. The time-segment sampling control module of the mirrored video stream extracts one frame every 300 milliseconds. After image enhancement and feature extraction processing, a unified feature vector containing multi-dimensional feature data is generated. The main data of this unified feature vector includes:

[0083] Road line features: 12 main lines were detected, with an average line length of 310 pixels, an average directional angle between 87° and 93°, and a directional distribution variance of 2.4.

[0084] Directional arrow features: Four arrow-shaped regions were identified, with an average arrow brightness value of 0.82 (normalized to the range of 0~1), and the average angle between the pointing direction and the main road lines was 8°;

[0085] Path trajectory characteristics: The average rate of change of path curvature is 0.03, the number of continuous trajectory segments is 5, and the average path width is 15 pixels;

[0086] Map symbol features: 18 text regions were identified, with an average spacing of 40 pixels and a brightness uniformity of 0.77;

[0087] Global brightness gradient: The overall brightness gradient is concentrated in the horizontal direction, accounting for approximately 84%;

[0088] Global texture distribution: The standard deviation of texture density is 0.12, indicating that the image texture is relatively uniform.

[0089] After inputting the above feature data into the intelligent discrimination model, the model's feature mapping layer first normalizes and performs non-linear mapping on these features to generate feature response vectors. These vectors are then input into the discrimination layer to determine whether the image belongs to the "navigation" or "non-navigation" category. After analyzing the features, the intelligent discrimination model outputs the following preliminary judgment results:

[0090] Navigation response value: 0.91;

[0091] Non-navigation response value: 0.09;

[0092] At this point, the model internally calculates the confidence weight for each category. The confidence weight is calculated based on the model's internal response strength to the degree of feature matching. Based on the mapping layer output, the model identifies feature combinations in the input image that highly match the navigation scenario, such as dense road lines, clear and stable directional arrows, continuous path trajectories, regular text distribution, and high texture uniformity. These features show strong consistency with typical navigation screens in the model's training data. By comparing the similarity between the input features and the internal feature template, the model obtains a confidence weight of 0.93 for the navigation category and 0.07 for the non-navigation category.

[0093] When calculating the confidence weights, the model weights each feature based on its matching degree. For example, the matching degree of road line direction stability with navigation features reaches 0.96, accounting for 30% of the total confidence weight calculation; the clear shape and consistent direction of the directional arrows have a matching degree of 0.94, accounting for 20%; the matching degree of path trajectory continuity is 0.91, accounting for 25%; the matching degree of map symbol distribution uniformity is 0.88, accounting for 15%; and the matching degree of overall brightness gradient consistency is 0.90, accounting for 10%. After weighting the above features, the comprehensive confidence weighted average is obtained as: (0.96×0.3) + (0.94×0.2) + (0.91×0.25) + (0.88×0.15) + (0.90×0.1) = 0.93, which is the comprehensive confidence weight of the navigation category.

[0094] To further illustrate the practical significance of this confidence weight, we can compare it to another scenario. If the vehicle is parked and the driver plays an entertainment video via mirror image, the feature extraction results in the image frame are as follows:

[0095] There is one road line, with a length of 80 pixels and a directional distribution variance of 45°.

[0096] No directional arrow feature;

[0097] No path trajectory features;

[0098] The text area has 3 regions, and the brightness uniformity is only 0.42;

[0099] The distribution of local brightness gradient directions is random, with no obvious concentration trend;

[0100] The standard deviation of texture density is 0.45, indicating that the texture changes drastically.

[0101] After the feature vector is input into the intelligent discrimination model, the model's output is as follows:

[0102] Navigation response value: 0.17;

[0103] Non-navigation response value: 0.83;

[0104] The corresponding confidence weights are: 0.19 for navigation and 0.81 for non-navigation. At this point, the feature matching accuracy within the model significantly decreases. Due to the lack of key features such as continuous lines, directional arrows, and path trajectories, the model automatically reduces the confidence weight of the navigation category based on the feature matching ratio.

[0105] The two examples above demonstrate that after identifying the unified feature vector input, the intelligent discrimination model not only provides a category judgment but also assigns a confidence weight to each category to reflect the model's confidence in the current judgment result. When the feature data contains a combination of elements that highly match navigation features, the model outputs a high confidence weight; conversely, when the features do not match the navigation screen features, the confidence weight decreases accordingly.

[0106] For example, in the first scenario mentioned above, the model outputs a confidence weight of 0.93 for navigation-related content, indicating that the system almost completely confirms that the current screen is a navigation interface. In the second scenario, however, the confidence weight for navigation-related content is only 0.19, while the confidence weight for non-navigation-related content is as high as 0.81, indicating that the system judges the current screen to be more likely to be video content. This confidence result will serve as the basis for subsequent dynamic correction and time-weighted calculations, thus forming a continuously updated high-confidence judgment output in the time dimension.

[0107] This specific example clearly demonstrates that after completing the initial identification, the intelligent discrimination model can calculate the confidence weight corresponding to each category based on the structural matching degree and feature distribution pattern of the input unified feature vector. This achieves the transformation from category identification to credible quantification, ensuring that the output contains both the judgment information of navigation and non-navigation categories and the quantitatively expressed confidence parameters, providing a reliable basis for subsequent dynamic correction and time series weight calculation.

[0108] After obtaining the recognition result including confidence weights, dynamic confidence correction is applied to the result to address continuous changes and feature fluctuations in the mirrored image over time. Specifically, the dynamic confidence correction strategy uses time as a reference to smoothly adjust the confidence weights between adjacent image frames, ensuring the stability of the recognition result across consecutive frames. This process is based on a weighted correction of the confidence difference between the previous and current frames. When the confidence change between consecutive frames is small, the system increases the weight retention of the current frame, allowing the recognition result to continue the judgment of the previous frame for a short period, thus avoiding erroneous switching caused by slight changes in the image. When the confidence change is large, it indicates that the image content may have changed, such as switching from a navigation interface to a video playback screen. In this case, the system increases the responsiveness of the current frame, allowing the recognition result to respond quickly to the change. In addition, considering the changes in lighting and screen reflection in the in-vehicle environment, the dynamic correction strategy incorporates a time decay factor during the weight adjustment process, allowing the confidence to gradually return to a neutral value when it remains unchanged for a long time, preventing outdated judgments from persisting for too long. By smoothing and decaying the confidence level over time, the identification results are made both stable and have real-time response capabilities, ensuring that the intelligent discrimination model maintains dynamic balance during continuous operation.

[0109] After dynamic confidence correction, the corrected results undergo time-weighted fusion processing to generate a high-confidence navigation screen discrimination output with time weights. Specifically, the system performs a weighted average of the confidence outputs of consecutive frames on the time axis, calculates the overall confidence trend based on the number of frames and weight distribution within the time window, and generates a comprehensive discrimination result including time weights. In this process, the allocation of time weights is determined based on the continuity between frames and the stability of content: if the confidence fluctuations of consecutive frames are small, it indicates that the screen content remains consistent in the time dimension, so the system increases the proportion of time weights, making the overall discrimination result more dependent on historical judgments; if consecutive frames change frequently, it indicates that there is dynamic switching of the screen, so the system reduces the proportion of time weights, increasing the immediate judgment influence of the current frame. Through this time-weighted fusion method, the recognition result can reflect the temporal stability of the navigation screen, while avoiding the impact of instantaneous interference on the judgment. Ultimately, the output not only includes the category determination of the navigation screen, but also the comprehensive confidence level after dynamic adjustment and time smoothing, making the entire discrimination result continuous in time, reasonable in logic, and controllable in stability, providing a high-confidence, sustainable and adaptive input basis for subsequent driving mode judgment.

[0110] Through the above steps, the intelligent discrimination model forms a complete dynamic discrimination mechanism encompassing feature input, category recognition, confidence correction, and time-weighted output. This mechanism not only accurately identifies the category of the navigation screen but also adjusts the judgment weights in real time according to changes in the screen, achieving a balance between temporal continuity and response sensitivity. The resulting high-confidence navigation screen discrimination output with time weights enables the in-vehicle terminal to maintain stable content recognition capabilities in complex driving scenarios, providing reliable data support and dynamic decision-making basis for driving mode judgment.

[0111] Step 5: Based on the recognition results with time weights output by the intelligent discrimination model, a time series decay and sliding window statistical mechanism is constructed to dynamically update the recognition status of multiple consecutive frames with weighted confidence. When the recognition results of consecutive frames are consistent, the system response update interval is extended, and when a sudden change in screen features is detected, the feedback cycle is shortened, thereby realizing adaptive release of navigation screens and closed-loop control of safety restrictions on non-navigation screens.

[0112] Based on the identification results with time weights output by the intelligent discrimination model, a time series decay and sliding window statistical mechanism is constructed. The specific steps are as follows:

[0113] The recognition results, weighted by time, output by the intelligent discrimination model are input into the time series management process, where they are sequentially arranged and time-calibrated on the timeline. Specifically, the timestamp, category label, and confidence weight associated with each frame's recognition result are recorded synchronously to form a continuous frame recognition sequence. This recognition sequence is constructed with a fixed time step, ensuring that each frame corresponds to a specific time interval. To guarantee the continuity of the time series, timestamps are aligned during frame sequence organization, ensuring a uniform distribution of time intervals for each recognition result, thus avoiding time drift issues caused by video transmission delays or unstable frame rates. During the establishment of the recognition sequence, the recognition result of the previous frame is used as a historical reference for the current frame, and the two are correlated through time weights, forming a continuous logical chain of recognition results in the time dimension. In this way, the recognition results of continuous frames not only maintain order on the timeline but also preserve the confidence influence relationship between consecutive frames, providing a structured input basis for subsequent sliding window statistics and confidence-weighted updates.

[0114] After constructing the time series, a sliding window statistical mechanism is used to dynamically analyze the recognition results of consecutive frames. Specifically, a sliding window of a preset fixed length is set on the time axis. This window contains several consecutive recognition frames, and the length of each window is dynamically matched according to the system refresh cycle and the frame rate of the mirrored video stream to ensure that the window contains a sufficient time span for state analysis during actual operation. When a new recognition frame is added to the sequence, the sliding window advances forward over time, removing the earliest recognition frame and including the latest frame in the analysis. For all recognition results within the window, statistical analysis is performed based on the confidence weight and time position of each frame to determine the stability and consistency of the recognition results over a continuous time period. When the recognition results within the window remain consistent in category and the confidence fluctuation is small, the system considers the current image state to be stable; conversely, when the results within the window show category switching or a sharp change in confidence, the image is considered to have undergone a sudden change. Through this sliding window statistical method, the system can perceive the trend of image changes in real time, providing a basis for subsequent dynamic confidence weighted updates.

[0115] After obtaining the statistical results of the sliding window, the recognition status within the window is dynamically updated with weighted confidence levels to balance the temporal stability and responsiveness of the recognition results. Specifically, the confidence levels of each frame within the sliding window are weighted according to their chronological order, giving higher weights to more recent frames and relatively lower weights to earlier frames, thus ensuring that the system's judgment is more sensitive to recent changes. While maintaining consistent recognition results, the system increases the overall weighted average based on the continuity of confidence levels, making the output results more stable and avoiding misjudgments caused by fluctuations in the confidence levels of individual frames. When the recognition results change, such as switching from a navigation screen to a video playback screen, the system automatically reduces the weight of historical frame confidence levels on the current result to quickly respond to the new screen state. Simultaneously, a confidence decay factor is introduced based on the rate of change of confidence in the time series. When the rate of change of confidence is low, the decay rate slows down to maintain the continuity of the recognition results; when the rate of change of confidence is high, the decay rate accelerates, enabling the system to quickly adjust its judgment direction. By using this time-series-based dynamic weighting method, the system can maintain flexible responsiveness while ensuring continuous identification, thus achieving a balance between stability and agility.

[0116] After completing the dynamic confidence-weighted update, the system adaptively adjusts the system response update interval and feedback cycle according to the changing trend of the time series to achieve control over the release of navigation screens and safety restriction control of non-navigation screens. Specifically, when the recognition results of consecutive frames remain consistent within the sliding window and the weighted confidence steadily increases, the system determines that the current screen state is stable and reliable. At this time, it automatically extends the recognition result update interval and reduces the feedback frequency to reduce system resource consumption and maintain a stable output state. When a sudden change in confidence, category switching, or drastic change in brightness structure is detected within the window, the system considers the screen to have changed abruptly, shortens the response interval to the minimum value, and increases the update frequency to achieve rapid response. For recognition results determined to be navigation screens, if the weighted confidence continuously exceeds a preset threshold, the system performs adaptive release of navigation screens, allowing them to be continuously displayed. For recognition results determined to be non-navigation content, the system immediately restores the safety restriction state, obscuring or pausing the display of video content to ensure driving safety. Through this adaptive adjustment method based on time series changes, the system can flexibly adjust the feedback cycle under different driving states, different ambient lighting conditions, and different changes in image content, so as to achieve a closed-loop control logic that can both ensure continuous display of navigation information and maintain driving safety.

[0117] Through the execution of the aforementioned consecutive steps, the time-series decay and sliding window statistical mechanism achieve dynamic correlation across three dimensions: time, confidence level, and response period. This enables the recognition process to automatically adjust the judgment frequency and release logic based on changes in the screen content. This mechanism not only ensures the continuity and visibility of the navigation screen during driving but also guarantees the restricted display of non-navigation screens under driving safety strategies, thereby achieving a balance between safety and intelligence. It provides a real-time, adaptive, and sustainable control strategy for driving mode judgment methods.

[0118] This invention introduces a computer vision-based content recognition mechanism into the in-vehicle terminal, enabling the system to automatically extract and analyze the visual features of the navigation screen from the mirrored video stream, achieving accurate recognition and dynamic release control of navigation-related screens. Through the coordinated processing of time-segment sampling and image enhancement, this solution allows the in-vehicle terminal to obtain high-quality, stable image input data even under limited computing power. This effectively avoids the problem of navigation screens being misinterpreted as video playback while driving, causing obstruction and interruption, thus ensuring continuous display of navigation information and driving safety.

[0119] This invention combines an intelligent discrimination model with a time-series sliding mechanism, enabling the recognition results to adaptively adjust in both spatial and temporal dimensions. The system can dynamically adjust its response frequency based on changes in the screen content, maintaining smooth output when recognition is stable and responding quickly to sudden changes in the screen. This achieves accurate differentiation and closed-loop control of navigation and non-navigation content, not only enhancing the intelligence and environmental adaptability of the in-vehicle system but also achieving an effective balance between driving safety and user experience.

[0120] This invention provides, for example Figure 4 The computer vision-based driving mode judgment system shown includes a time segment sampling control module, an image enhancement processing module, a feature extraction and analysis module, an intelligent discrimination and recognition module, and a time series dynamic update module.

[0121] The time segment sampling control module constructs a time segment sampling control mechanism based on the mirrored video stream received by the vehicle system, extracts static image frames at fixed time intervals, and dynamically adjusts the frame extraction frequency according to the resource load status of the vehicle system.

[0122] The image enhancement processing module inputs the static image frames obtained by sampling time segments into the image enhancement processing flow, performs local contrast adaptive adjustment on the static image frames, enhances the geometric features of navigation elements, and obtains an enhanced image with balanced visual features.

[0123] The feature extraction and analysis module will enhance the image input feature extraction process. Based on multi-level visual feature analysis, it will perform local structure recognition and global texture parsing, extract multi-dimensional visual feature information, and fuse brightness gradient and texture distribution information to generate a unified feature vector.

[0124] The intelligent discrimination and recognition module inputs a unified feature vector into the intelligent discrimination model, performs category recognition on the navigation screen through a nonlinear feature mapping relationship, outputs a recognition result containing confidence weights, and performs weighted adjustment on the recognition result based on a dynamic confidence correction strategy;

[0125] The time series dynamic update module constructs a time series decay and sliding window statistical mechanism based on the recognition results with time weights output by the intelligent discrimination model, and performs dynamic confidence weighted updates on the recognition status of multiple consecutive frames.

[0126] The computer vision-based driving mode determination method provided in this embodiment of the invention is implemented through the aforementioned computer vision-based driving mode determination system. For details of the specific methods and processes of the computer vision-based driving mode determination system, please refer to the embodiments of the aforementioned computer vision-based driving mode determination method, which will not be repeated here.

[0127] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A driving mode determination method based on computer vision, characterized in that, Includes the following steps: Step 1: Based on the mirrored video stream received by the vehicle system, a time segment sampling control mechanism is constructed to extract static image frames at fixed time intervals and dynamically adjust the frame extraction frequency according to the resource load status of the vehicle system. Step 2: Input the static image frames obtained from time segment sampling into the image enhancement processing flow, perform local contrast adaptive adjustment on the static image frames, and obtain the enhanced image; Step 3 enhances the image input feature extraction process by performing local structure recognition and global texture parsing based on multi-level visual feature analysis, extracting multi-dimensional visual feature information, and fusing brightness gradient and texture distribution information to generate a unified feature vector. Step 4: Input the unified feature vector into the intelligent discrimination model, perform category recognition on the navigation screen through nonlinear feature mapping relationship, output the recognition result containing confidence weight, and adjust the recognition result based on dynamic confidence correction strategy. Step 5: Based on the recognition results with time weights output by the intelligent discrimination model, construct a time series decay and sliding window statistical mechanism to dynamically update the confidence weight of the recognition status of multiple consecutive frames. The unified feature vector is input into the intelligent discrimination model, and the navigation screen is classified through nonlinear feature mapping relationship. The specific steps are as follows: The generated unified feature vector is input into the intelligent discrimination model to perform initial classification and recognition. The correspondence between features and navigation categories is established through feature normalization and multi-layer nonlinear mapping to obtain preliminary classification results. After the initial identification by the intelligent discrimination model is completed, the confidence level and weight allocation are performed on the initial classification results output by the intelligent discrimination model. The confidence level weights of each category are determined based on the discrimination strength and feature matching degree within the intelligent discrimination model. After obtaining the recognition result including confidence weights, the recognition result is dynamically confidence corrected, and the confidence weights are smoothly adjusted between adjacent image frames with time as a reference. After completing the dynamic confidence correction, the corrected recognition results are subjected to time-weighted fusion processing, and a recognition result with time weight is generated by weighted averaging within the time window. Based on the identification results with time weights output by the intelligent discrimination model, the specific steps for constructing the time series decay and sliding window statistical mechanism are as follows: The recognition results with time weights output by the intelligent discrimination model are input into the time series management process. The recognition results are arranged sequentially and time-calibrated on the time axis. The timestamp, category label and confidence weight of each frame recognition result are recorded synchronously. After the time series is constructed, the sliding window statistical mechanism is used to dynamically analyze the recognition results of consecutive frames. A sliding window of a preset fixed length is set on the time axis, and the confidence weight and time position of the recognition results of each frame within the sliding window are statistically judged. After obtaining the statistical results of the sliding window, the recognition status within the sliding window is dynamically updated with weighted confidence. The confidence of each frame is weighted according to the time sequence, and a confidence decay factor is introduced based on the confidence change rate. After completing the dynamic confidence weighted update, the system response update interval and feedback cycle are adaptively adjusted according to the time series change trend to achieve the release control of navigation screen and the safety restriction control of non-navigation screen. When the recognition results of consecutive frames are consistent, the system response update interval is extended, and when a sudden change in screen features is detected, the feedback cycle is shortened.

2. The driving mode determination method based on computer vision according to claim 1, characterized in that, The specific steps for constructing a time-segment sampling control mechanism based on the mirrored video stream received by the vehicle system are as follows: After the vehicle terminal receives the mirrored video stream from the mobile device, it segments the mirrored video stream according to the timeline. By reading the timestamp information of the video frames, the video stream is divided into multiple consecutive time segments, and the video frames in each time segment are numbered and registered to generate a frame index table. After the time segments are divided, static image frames are extracted from each time segment at fixed time intervals, and the frame extraction frequency is dynamically adjusted according to the resource load status of the vehicle terminal. After dynamic frame extraction is completed, the extracted static image frames are organized and cached sequentially. The logical positions between frames are corrected through timestamp alignment, and a cache queue is established according to the first-in-first-out principle.

3. The computer vision-based driving mode determination method according to claim 1, characterized in that, The static image frames obtained from time segment sampling are input into the image enhancement processing flow. The specific steps for performing local contrast adaptive adjustment on the static image frames are as follows: The static image frames obtained by time segment sampling are input into the image enhancement processing flow, and basic normalization processing is performed on each static image frame; After completing the basic standardization process, local region segmentation and partition brightness analysis are performed on the adjusted static image frames. Based on the brightness distribution range and contrast change characteristics of each region, differentiated brightness adjustment is performed to enhance the brightness difference in regions containing road elements. After completing the local brightness partition adjustment, geometric enhancement and edge smoothing optimization are performed on the adjusted static image frames. By identifying the brightness gradient direction, the boundaries of the main road and the path trajectory are enhanced, and the overlapping area is smoothed to obtain the enhanced image frame. After completing the geometric enhancement, global visual equalization and continuous frame brightness smoothing are performed. The enhanced image frames are then redistributed and their continuity is corrected in both spatial and temporal dimensions to generate a continuously enhanced image sequence.

4. The computer vision-based driving mode determination method according to claim 3, characterized in that, In the process of global visual equalization and continuous frame brightness smoothing, the average brightness curve of the enhanced image frame is dynamically aligned, and the brightness transition between adjacent frames is adjusted based on the time sequence of brightness distribution of each frame.

5. The computer vision-based driving mode determination method according to claim 3, characterized in that, The image input feature extraction process will be enhanced to generate a unified feature vector for content discrimination. The specific steps are as follows: The enhanced image input feature extraction process divides the spatial distribution of the enhanced image and identifies key regions. By analyzing the brightness distribution, shape boundaries, and pixel arrangement direction, regions with structural regularities are identified, and their spatial location, shape contours, and connection relationships are recorded. After obtaining the spatial distribution division and key area identification results, geometric structure analysis is performed on each local area to extract the geometric features of the road area, and the extracted geometric features are spatially indexed and arranged to form local structural features. After extracting local structural features, global visual feature analysis is performed on the enhanced image. By statistically analyzing the brightness gradient direction, color distribution, and texture density, global visual features reflecting the overall layout features of the navigation screen are formed, and the global visual features are aligned with the local structural features in spatial coordinates. Local structural features and global visual features are fused and integrated into a unified feature vector according to spatial location and content weight.

6. The computer vision-based driving mode determination method according to claim 5, characterized in that, When fusing local structural features with global visual features, the direction information of road lines is matched with the direction of global brightness distribution based on the spatial position correspondence, the path contour morphology parameters are associated with global texture density, and the direction arrow angle information is combined with the brightness concentration trend.

7. The driving mode determination method based on computer vision according to claim 1, characterized in that, The time-weighted fusion processing includes: dynamically adjusting the time weight allocation based on the confidence fluctuation of the recognition results of consecutive frames within the time window; increasing the weight of historical frames when the recognition results remain stable; and increasing the weight of the current frame when the recognition results change.

8. A computer vision-based driving mode determination system, used to implement the computer vision-based driving mode determination method according to any one of claims 1-7, characterized in that, It includes a time segment sampling control module, an image enhancement processing module, a feature extraction and analysis module, an intelligent discrimination and recognition module, and a time series dynamic update module; The time segment sampling control module constructs a time segment sampling control mechanism based on the mirrored video stream received by the vehicle system, extracts static image frames at fixed time intervals, and dynamically adjusts the frame extraction frequency according to the resource load status of the vehicle system. The image enhancement processing module inputs the static image frames obtained by sampling time segments into the image enhancement processing flow, performs local contrast adaptive adjustment on the static image frames, and obtains the enhanced image. The feature extraction and analysis module will enhance the image input feature extraction process. Based on multi-level visual feature analysis, it will perform local structure recognition and global texture parsing, extract multi-dimensional visual feature information, and fuse brightness gradient and texture distribution information to generate a unified feature vector. The intelligent discrimination and recognition module inputs a unified feature vector into the intelligent discrimination model, performs category recognition on the navigation screen through a nonlinear feature mapping relationship, outputs a recognition result containing confidence weights, and performs weighted adjustment on the recognition result based on a dynamic confidence correction strategy; The time series dynamic update module constructs a time series decay and sliding window statistical mechanism based on the recognition results with time weights output by the intelligent discrimination model, and performs dynamic confidence weighted updates on the recognition status of multiple consecutive frames.

Images