Complex logistics scene-oriented adaptive contrast enhancement video fusion algorithm

Abstract

The invention provides an adaptive contrast enhancement video fusion algorithm for a complex logistics scene, and the algorithm comprises the steps: extracting environment features from an initial video data set, analyzing the inter-frame illumination change and object movement speed through employing a convolutional neural network, determining that the current scene is a peak period or a low-illumination environment, and obtaining environment change perception parameters; extracting a key frame from the second video data set, separating a foreground object and a background by adopting an image segmentation technology based on deep learning, and performing detail retention enhancement on the foreground object to obtain a third video data set; real-time verification is carried out on the final video output, if the processing delay exceeds a preset threshold value, the resource distribution proportion is adjusted through a feedback mechanism, mode switching and enhancement processing are executed again, and the updated video output is obtained.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an adaptive contrast-enhanced video fusion algorithm for complex logistics scenarios. Background Technology

[0002] Background of the problem: Video processing technology is crucial for improving operational efficiency and accuracy in logistics scenarios, especially in complex warehousing, sorting, and transportation environments where high-definition, real-time video data can significantly optimize decision-making and monitoring processes. However, existing video enhancement and fusion methods often exhibit significant limitations when facing the diversity of logistics scenarios. Many traditional algorithms rely on fixed processing patterns, making it difficult to adapt to dynamically changing scenario requirements. For example, they may experience insufficient processing speed during peak periods or poor enhancement effects in low-light environments. These shortcomings prevent logistics systems from providing reliable video support at critical moments, thereby impacting overall efficiency.

[0003] Complex logistics scenarios present unique technical challenges to video processing algorithms. Video data exhibits significantly different characteristics across different time periods and environments, such as massive data streams during peak hours versus low-light conditions at night, requiring algorithms to dynamically adjust. Fixed processing strategies struggle to balance speed and quality, leading to processing delays under high loads or loss of image details in complex environments. Dynamic adjustment relies on the algorithm's intelligent allocation of computing resources, but the optimization of resource allocation is constrained by real-time requirements. Increased real-time demands further require algorithms to rapidly switch processing modes with limited computing resources; however, the frequency and stability of mode switching are interdependent, increasing the complexity of algorithm design.

[0004] Therefore, the key issue of this research is how to design a video fusion algorithm that can dynamically adjust the processing strategy according to the needs of the scenario and the computing resources in complex logistics scenarios, so as to achieve the best balance between speed, quality and stability. Summary of the Invention

[0005] This invention provides an adaptive contrast-enhanced video fusion algorithm for complex logistics scenarios, mainly comprising: Acquire video data streams of logistics scenarios, capture multi-channel video signals through a preset sensor array, and mark timestamps and light intensity for peak data streams and low-light environments respectively to obtain an initial video dataset; Environmental features are extracted from the initial video dataset. Convolutional neural networks are used to analyze inter-frame lighting changes and object movement speed to determine whether the current scene is a peak period or a low-light environment, thus obtaining environmental change perception parameters. For the resource allocation scheme, the corresponding video processing mode is activated, and the frame rate and resolution are adjusted through the preset mode switching algorithm. The gain parameter is increased for low light environment, and the resolution is reduced for peak period to improve the processing speed, so as to obtain the optimized video processing parameters. The initial video dataset is enhanced using optimized video processing parameters. Noise in low-light environments is removed using an adaptive denoising algorithm, and data redundancy during peak periods is reduced using a fast compression algorithm to obtain the second video dataset. Keyframes were extracted from the second video dataset, and deep learning-based image segmentation techniques were used to separate foreground objects from the background. Detail preservation enhancement was performed on the foreground objects to obtain the third video dataset. A fused video stream is generated based on the third video dataset. Multi-channel video signals are integrated through a spatiotemporal fusion algorithm. Inter-frame interpolation is optimized for peak data streams, and contrast is enhanced for low-light environments to obtain the final video output. Real-time verification is performed on the final video output. If the processing delay exceeds the preset threshold, the resource allocation ratio is adjusted through the feedback mechanism, and the mode switching and enhancement processing are re-executed to obtain the updated video output. The updated video output is used to predict the scene change trend in the next minute using a recurrent neural network. The preset threshold and mode switching frequency are adjusted to obtain adaptive adjustment parameters for the next round of video data stream processing.

[0006] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses an adaptive contrast-enhanced video fusion algorithm for complex logistics scenarios, specifically designed for processing logistics video data streams during peak hours and in low-light environments. The invention first acquires video signals using multi-channel sensors, extracts environmental features, and detects scene changes. Then, it adaptively adjusts video processing parameters according to different scenarios, such as increasing gain in low-light environments and reducing resolution during peak hours. Next, it performs enhancement processing on the video data, including denoising and compression, and utilizes deep learning technology to segment foreground objects for detail preservation. Finally, it integrates multi-channel video using a spatiotemporal fusion algorithm, specifically optimizing inter-frame interpolation and contrast to generate high-quality video output. This invention also features real-time verification and adaptive adjustment capabilities, dynamically adjusting resource allocation based on processing latency and predicting scene change trends to optimize processing parameters, thereby achieving efficient video processing and fusion in complex logistics scenarios. Attached Figure Description

[0007] Figure 1 This is a flowchart of an adaptive contrast-enhanced video fusion algorithm for complex logistics scenarios according to the present invention.

[0008] Figure 2This is a schematic diagram of an adaptive contrast-enhanced video fusion algorithm for complex logistics scenarios according to the present invention.

[0009] Figure 3 This is another schematic diagram of an adaptive contrast-enhanced video fusion algorithm for complex logistics scenarios according to the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0011] like Figure 1-3 This embodiment of an adaptive contrast enhancement video fusion algorithm for complex logistics scenarios may specifically include: Step S101: Obtain video data stream of logistics scene. Capture multi-channel video signals through a preset sensor array. Mark the timestamp and light intensity for peak period data stream and low light environment respectively to obtain the initial video dataset.

[0012] An initial video dataset is obtained by capturing multi-channel video signals of a logistics scene using a pre-set sensor array, and labeling them with timestamps and illumination intensity. If the initial video dataset contains peak-period data, dynamic objects are detected using inter-frame difference to obtain a sequence of dynamic objects. Based on the sequence of dynamic objects, optical flow is used to calculate the motion trajectories of the objects, resulting in a set of motion trajectories. If the trajectory density in the motion trajectory set exceeds a preset threshold, a clustering algorithm is used to divide congested areas, resulting in a distribution of congested areas. Based on the distribution of congested areas, video frames under low-light conditions are acquired, and histogram equalization is used to enhance image contrast, resulting in an enhanced image set. Keyframe features are extracted from the enhanced image set, and a pre-set template matching method is used to identify logistics targets, resulting in target recognition results. Based on the target recognition results, timestamps and illumination intensity are fused to generate a logistics scene status dataset.

[0013] Specifically, in the acquisition of video data streams in logistics scenarios, a sensor array consisting of eight 4K resolution cameras is first deployed to synchronously acquire multi-channel video signals at a sampling rate of 30 frames per second. Each camera is equipped with a Sony IMX586 sensor, supporting low-light imaging down to 0.1 lux. For peak-period data streams, a dynamic frame rate adjustment algorithm based on a time window is adopted. When the number of pedestrians detected exceeds 150 per minute, the algorithm automatically switches to 60 frames per second mode and synchronously records timestamps accurate to the millisecond level via the NTP protocol. In low-light environments, ambient illuminance is monitored in real time by a photosensor. When the illuminance value is below 10 lux, an infrared supplementary lighting module is triggered. Simultaneously, OpenCV's CLAHE algorithm (clip limit=2.0, tile grid size=8x8) is used for image enhancement, and the illuminance sensor readings are hard-synchronized with the video frames. In the initial dataset construction phase, the YOLOv5s model was used for multi-target detection of people, vehicles, and objects, with a confidence threshold set at 0.7. The Hungarian algorithm was used for cross-frame tracking of the detection results, ultimately generating structured metadata containing timestamps, illumination intensity, target category, and motion trajectory. In the data preprocessing stage, bicubic interpolation was used to uniformly scale the resolution of data acquired from different cameras to 1920x1080, and the KLT optical flow method was used to calculate the motion vectors between adjacent frames. Frames with motion amplitudes exceeding 15 pixels were marked as keyframes. To eliminate sensor differences, intrinsic parameter matrices were calculated for each camera based on the Zhang Zhengyou calibration method, and extrinsic parameter calibration used the PnP algorithm, with reprojection errors controlled within 1.2 pixels. The final generated initial dataset includes time-aligned multi-view video streams, frame-by-frame metadata, and sensor calibration parameters, providing complete input for subsequent analysis.

[0014] Step S102: Extract environmental features from the initial video dataset, use a convolutional neural network to analyze inter-frame illumination changes and object movement speed, determine whether the current scene is a peak period or a low-light environment, and obtain environmental change perception parameters.

[0015] Frame sequences are extracted from a video dataset, and inter-frame differencing is applied to calculate illumination differences and object motion trajectories, yielding illumination change features and motion speed features. A convolutional neural network (CNN) is used to process the illumination change features, extracting the spatial distribution pattern of illumination between frames to obtain illumination distribution parameters. A CNN is also used to process the motion speed features, analyzing the temporal change pattern of object motion to obtain motion distribution parameters. If the illumination distribution parameters are below a preset threshold, the environment is identified as low-light, generating low-light perception parameters. If the motion distribution parameters are above a preset threshold, the scene is identified as a peak-period scene, generating peak-period perception parameters. Based on the low-light and peak-period perception parameters, the environmental perception results are fused to generate comprehensive environmental perception parameters. Using these comprehensive environmental perception parameters, the scene judgment model is updated, and the feature extraction of illumination differences and motion speed is optimized to obtain adaptive environmental perception parameters.

[0016] Specifically, when extracting environmental features from the initial video dataset, the inter-frame difference method is first used to calculate the pixel changes between adjacent frames. A threshold of 30 is set to detect areas of significant change. For example, in traffic monitoring videos, when a vehicle's speed exceeds 5 meters per second, the inter-frame difference results will show a clear motion trajectory. Next, a convolutional neural network (CNN) is used to analyze the video frames. The network structure uses ResNet-50, with an input of a 224x224 RGB image. Illumination features are extracted through convolutional layers. For example, in low-light environments, the average brightness value of the image may be below 50, while during peak periods with sufficient light, the average brightness value may exceed 150. To further analyze the object's moving speed, optical flow is used to calculate the motion vector of pixels. For example, in traffic scenes, the average optical flow speed of vehicles may reach 10 pixels per second, while in pedestrian scenes, the optical flow speed may only be 2 pixels per second. By combining changes in illumination and object movement speed, a Support Vector Machine (SVM) classifier is used for scene classification. For example, when the illumination value is below 60 and the object movement speed is above 8 pixels per second, it is classified as a peak scene; when the illumination value is below 40 and the object movement speed is below 3 pixels per second, it is classified as a low-light environment. Finally, the system outputs environmental change perception parameters, including illumination intensity, object movement speed, and scene type. For example, if the illumination intensity is 45, the object movement speed is 6 pixels per second, and the scene type is peak, these parameters can be used for subsequent optimization of the intelligent monitoring system, such as adjusting the camera's exposure time and frame rate to adapt to different environmental conditions.

[0017] Step S103: For the resource allocation scheme, activate the corresponding video processing mode, adjust the frame rate and resolution through the preset mode switching algorithm, increase the gain parameter for low light environment, and reduce the resolution for peak period to improve processing speed, so as to obtain the optimized video processing parameters.

[0018] The system acquires current light intensity data using an ambient light sensor and uses a preset threshold to determine the lighting environment type, resulting in a classification of low light or normal light. If the classification result is low light, a preset gain adjustment algorithm is used to increase the gain parameter, generating the first video processing parameter. The system then acquires current load data and uses a preset load threshold to determine if it is a peak period, resulting in a peak or off-peak period judgment. If the judgment result is a peak period, a resolution reduction algorithm is used to adjust the resolution, generating the second video processing parameter. Based on the first and second video processing parameters, a parameter fusion algorithm is used to generate comprehensive video processing parameters. A frame rate adjustment module switches the frame rate of the comprehensive video processing parameters to obtain optimized video parameters.

[0019] Specifically, in the resource allocation scheme, the video processing parameters are first dynamically adjusted through a mode switching algorithm. For example, when the system detects that the current frame rate is 30fps and the resolution is 1080p, if the computational resource utilization rate exceeds 80%, the resolution reduction logic is triggered, reducing the resolution to 720p while maintaining the frame rate. At this time, a bilinear interpolation algorithm is used for rapid scaling to ensure that the processing latency is controlled within 50ms. For low-light environments, the system automatically adjusts the gain parameters based on image signal-to-noise ratio (SNR) analysis. When the ambient illuminance is below 50 lux, a 3dB gain boost is initiated, and combined with temporal filtering (window size of 5 frames) in the 3D noise reduction algorithm, noise is suppressed, so that the peak signal-to-noise ratio (PSNR) of the output image is maintained above 30dB. During peak traffic scenarios, an adaptive strategy based on queue length is adopted. When the number of frames to be processed exceeds 200, a fast processing mode is activated, further reducing the resolution to 480p and enabling the H264 fast encoding preset (--preset ultrafast), reducing the single-frame encoding time from 15ms to 8ms. Simultaneously, optical flow is used to compensate for inter-frame motion to maintain subjective smoothness. All parameter adjustments are implemented using a weighted decision tree algorithm, with resource utilization weighted at 0.6, latency weighted at 0.3, and image quality weighted at 0.1. When the overall score falls below the threshold of 0.7, a mode switch is triggered, and the switching process uses a cosine gradient curve transition to avoid abrupt changes in the image.

[0020] Step S104: The initial video dataset is enhanced using optimized video processing parameters. Noise in low-light environments is removed using an adaptive denoising algorithm, and data redundancy during peak periods is reduced using a fast compression algorithm, resulting in the second video dataset.

[0021] An adaptive denoising algorithm is used to process the initial video dataset. For noise in low-light environments, a weighted average of the grayscale value of a pixel and its neighboring pixels is calculated. If the grayscale value deviation exceeds a preset threshold, it is replaced with the weighted average, resulting in a denoised first intermediate video dataset. A fast compression algorithm is then used to process the first intermediate video dataset, analyzing the inter-frame pixel change rate. If the change rate is below a preset threshold, similar frames are merged to reduce data redundancy, resulting in a compressed second intermediate video dataset. Based on the inter-frame features of the second intermediate video dataset, edge information of keyframes is extracted, and an edge intensity map is generated using an edge detection algorithm, resulting in a third intermediate video dataset containing edge features. An adaptive sharpening algorithm is applied to the edge features of the third intermediate video dataset to enhance the edge contrast of keyframes. If the edge intensity is below a preset threshold, the sharpening weight is increased, resulting in a sharpened fourth intermediate video dataset. The brightness distribution of the fourth intermediate video dataset is obtained, and the pixel values ​​in low-light areas are analyzed. If the pixel values ​​are below a preset brightness threshold, gamma correction is used to adjust the brightness, resulting in a brighter fifth intermediate video dataset. By using the frame sequence of the fifth intermediate video dataset, inter-frame motion vectors are detected, and a motion compensation algorithm is used to optimize the smoothness of the inter-frames. If the motion vector deviation exceeds a preset threshold, a compensation frame is inserted to obtain a smooth second video dataset.

[0022] Specifically, in the video processing, the initial video dataset is first enhanced using optimized parameters. This includes increasing the frame rate from 30fps to 60fps, upgrading the resolution from 1080p to 4K, and enhancing the color dynamic range using HDR technology, resulting in richer video details. Next, an adaptive denoising algorithm is used to remove noise in low-light environments. This algorithm, based on the BM3D (Block-Matching and 3D Filtering) model, analyzes the noise distribution in video frames and automatically adjusts the denoising intensity. For example, in scenes with lighting below 50 lux, the noise level is reduced from 30dB to 15dB, significantly improving image clarity. Subsequently, a fast compression algorithm is used to reduce data redundancy during peak periods. This algorithm, based on the HEVC (High Efficiency Video Coding) standard, uses inter-frame prediction and transform coding techniques to compress the video bitrate from 10Mbps to 4Mbps while maintaining a PSNR (Peak Signal-to-Noise Ratio) above 40dB, ensuring no significant loss of video quality. Finally, after the above processing, a second video dataset is obtained. This dataset maintains high quality while significantly reducing storage and transmission costs, making it suitable for large-scale video analysis and application scenarios.

[0023] Step S105: Extract keyframes from the second video dataset, use deep learning-based image segmentation technology to separate foreground objects and background, perform detail preservation enhancement on foreground objects, and obtain the third video dataset.

[0024] Frame sequences were obtained from the second video dataset. A pre-trained deep convolutional neural network model was used to extract keyframes through inter-frame difference analysis, resulting in a keyframe set. Semantic segmentation algorithms were applied to each frame in the keyframe set, separating foreground objects and background based on a deep learning model to generate foreground and background masks. For the foreground mask region, a super-resolution algorithm was applied to enhance details, generating an enhanced foreground image. If the pixel value of the foreground mask region was below a preset threshold, texture restoration was performed on that region, resulting in a restored foreground image. The restored foreground image was then fused with the background mask region to generate an enhanced complete frame image, resulting in an enhanced frame set. Frame sequences were obtained from the enhanced frame set and reassembled according to timestamp order to generate the third video dataset. Based on the inter-frame consistency of the third video dataset, an optical flow algorithm was used to optimize inter-frame transitions, resulting in a smooth third video dataset.

[0025] Specifically, when extracting keyframes from the second video dataset, a method based on inter-frame differences can be used. A threshold of 0.3 is set, and the frames with differences exceeding the threshold are selected as keyframes by calculating the pixel differences between consecutive frames.

[0026] For example, the `absdiff` function from the OpenCV library is used to calculate inter-frame differences, combined with Gaussian filtering to remove noise, ultimately extracting keyframes. For the extracted keyframes, deep learning-based image segmentation techniques, such as the DeepLabV3+ model, are employed to separate foreground objects from the background. This model uses ResNet-101 as the backbone network, with an input image resolution of 512×512, and outputs a class probability distribution for each pixel. By setting a confidence threshold of 0.8, pixels with probabilities greater than this value are classified as foreground objects, and the rest as background. To enhance the details of the foreground objects, a generative adversarial network (GAN)-based enhancement method can be used, such as the ESRGAN model. This model generates high-resolution images through adversarial training, with the segmented foreground objects as input and the enhanced images as output. Specific parameters are set as follows: generator learning rate 0.0001, discriminator learning rate 0.0004, and 1000 training iterations. Finally, the enhanced foreground objects are re-synthesized with the background to obtain the third video dataset. The entire process is implemented through automated scripts to ensure the continuity and efficiency of data processing.

[0027] Step S106: Generate a fused video stream based on the third video dataset, integrate multi-channel video signals through a spatiotemporal fusion algorithm, optimize inter-frame interpolation for peak data streams, enhance contrast for low-light environments, and obtain the final video output.

[0028] Raw data is acquired from multi-channel video signals and integrated using a spatiotemporal fusion algorithm to generate a fused video stream. Inter-frame interpolation is used to process the fused video stream, optimizing it for peak data streams to obtain a smooth video sequence. If the smooth video sequence is in a low-light environment, a contrast enhancement algorithm is used to adjust brightness and contrast to obtain an enhanced video sequence. Environmental adaptability detection is performed to determine if the enhanced video sequence meets output quality requirements; if not, contrast parameters are iteratively adjusted to obtain an optimized video sequence. Based on the optimized video sequence, keyframe data is acquired, and data stream processing techniques are used to compress redundant information to generate a compressed video stream. For the compressed video stream, output quality enhancement processing is performed, removing noise using a denoising algorithm to obtain the final video output. Feature data is extracted from the final video output and stored in a preset database for subsequent signal integration.

[0029] Specifically, in the process of generating the fused video stream, multi-channel video signals are first integrated using a spatiotemporal fusion algorithm. For example, a motion compensation algorithm based on optical flow is used to align the 1080p@30fps video streams captured by three cameras. The optical flow calculation uses the Farneback algorithm, with the polynomial expansion order set to 2 and the optical flow window size set to 15×15 pixels. The motion-compensated frames are then fused by weighted averaging, with the weight coefficients dynamically adjusted according to the signal-to-noise ratio (SNR). When the SNR is higher than 30dB, the weight of the main channel is set to 0.7, and the weight of each auxiliary channel is 0.15. For peak data stream optimization, when frame rate fluctuations exceed ±5fps, an adaptive inter-frame interpolation algorithm is enabled. This algorithm uses the SuperSlomo model based on a convolutional neural network, taking two consecutive frames of images and an optical flow graph as input and outputting the intermediate interpolated frame. On a Tesla V100 graphics card, a single inference takes 23ms, ensuring a stable output of 30fps. For low-light environments, when the average image brightness is detected to be below 50 lux, a hybrid enhancement process is initiated: first, the CLAHE algorithm is used for contrast stretching with a block size of 8×8 and a histogram cropping limit of 2.0; then, illumination correction is performed using a RetinexNet network with the network input size adjusted to 512×512. With NPU acceleration, the processing latency is controlled within 40ms, resulting in a PSNR value exceeding 28.5dB and an SSIM of 0.92 for the final output video. The entire processing pipeline adopts a pipelined parallel architecture, with each module achieving zero-wait switching through a circular buffer, and latency jitter controlled within ±2ms.

[0030] Step S107: Perform real-time verification on the final video output. If the processing delay exceeds the preset threshold, adjust the resource allocation ratio through the feedback mechanism, re-execute the mode switching and enhancement processing, and obtain the updated video output.

[0031] Real-time video data is acquired from the video stream, and the processing latency is calculated using the inter-frame difference method to obtain the latency value. If the latency value exceeds a preset threshold, a resource adjustment signal is generated through a preset feedback mechanism to determine the resource allocation ratio. Based on the resource adjustment signal, a dynamic scheduling algorithm is used to reallocate computing resources, resulting in an updated resource allocation scheme. For the updated resource allocation scheme, a mode switching operation is performed to generate the switched processing mode. Using the switched processing mode, an image enhancement algorithm is used to process the video data, resulting in enhanced video frames. Output data is extracted from the enhanced video frames to generate the updated video output. If the latency of the updated video output still exceeds the preset threshold, the process returns to the resource adjustment signal generation step, repeating until the latency meets the threshold, resulting in the final video output.

[0032] Specifically, in the real-time video processing system, a timestamp-based latency detection algorithm is first deployed. For example, a sliding window is used to calculate the average processing latency of the most recent 5 frames. When the detected latency exceeds a 50-millisecond threshold, a feedback mechanism is triggered. The system adjusts computing resources through a dynamic resource allocation algorithm (such as a PID-based resource regulator), increasing GPU utilization from 60% to 80% while reducing the number of CPU decoding threads from 8 to 4. The mode switching module uses a decision tree algorithm to automatically switch from full-resolution processing to a 3840×2160@30fps subsampling mode based on the current latency value (e.g., 65 milliseconds) and content complexity (determined as a high-complexity scene by a DCT coefficient variance > 120). The enhancement processing stage uses an adaptive algorithm. When the network bandwidth is detected to be below 15Mbps, the QP value of the H265 encoding is adjusted from 22 to 28, and the detail enhancement module for non-key frames is disabled. The updated video stream undergoes a secondary quality assessment, using the VMAQ algorithm to ensure a score of at least 85. If this score is not met, reinforcement learning-based parameter optimization is triggered, adjusting the bitrate allocation weight coefficient β from 0.3 to 0.45 iteratively. The entire closed-loop system continuously monitors the system at 200-millisecond intervals. When a Kalman filter predicts a potential 72-millisecond delay in the next cycle, a resource pre-allocation strategy is initiated in advance.

[0033] Step S108: Using the updated video output, a recurrent neural network is used to predict the scene change trend within the next minute, and the preset threshold and mode switching frequency are adjusted to obtain adaptive adjustment parameters for the next round of video data stream processing.

[0034] Real-time scene frame sequences are acquired from video stream data, and dynamic features are extracted using the frame difference method to obtain a scene change feature set. For this feature set, a recurrent neural network (RNN) is used for time series analysis to generate a scene change trend for the next minute. If the scene change trend exceeds a preset threshold, the threshold is adjusted using gradient descent to obtain updated threshold parameters. Based on the updated threshold parameters, the mode switching frequency is calculated to generate an adaptive switching frequency value. The video stream processing mode is adjusted using the adaptive switching frequency value to obtain an optimized processing mode configuration. Subsequent video stream data is processed using the optimized processing mode configuration to generate real-time processing results. If the real-time processing results deviate significantly from the scene change trend, the RNN parameters are updated through a feedback mechanism to obtain an improved prediction model.

[0035] Specifically, in video data processing, the current video frame sequence is first modeled using a recurrent neural network (RNN), with a long short-term memory network (LSTM) as the core algorithm. The input is video data at 30 frames per second, and the time step is set to 60 seconds, or 1800 frames. The RNN learns the temporal relationships between video frames through training and predicts the scene change trend within the next minute.

[0036] For example, in a surveillance scenario, an RNN can predict the direction of pedestrian movement or changes in vehicle speed, outputting predictions for each frame over the next 60 seconds. Then, based on the predictions, preset thresholds are dynamically adjusted; for instance, the motion detection threshold is adjusted from an initial 0.7 to 0.5 to accommodate a decrease in moving targets in the scene. Simultaneously, the mode switching frequency is adjusted from once per second to once every 2 seconds to reduce computational overhead. These adaptive adjustment parameters are updated in real-time through a feedback mechanism and applied to the next round of video data stream processing.

[0037] For example, in the next round of processing, the system will re-detect moving targets based on new thresholds and frequencies, and optimize resource allocation. In this way, the system can continuously adapt to changes in the scene, improving processing efficiency and accuracy.

[0038] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive contrast-enhanced video fusion algorithm for complex logistics scenarios, characterized in that, The method includes: Acquire video data streams of logistics scenarios, capture multi-channel video signals through a preset sensor array, and mark timestamps and light intensity for peak data streams and low-light environments respectively to obtain an initial video dataset; Environmental features are extracted from the initial video dataset. Convolutional neural networks are used to analyze inter-frame lighting changes and object movement speed to determine whether the current scene is a peak period or a low-light environment, thus obtaining environmental change perception parameters. For the resource allocation scheme, the corresponding video processing mode is activated, and the frame rate and resolution are adjusted through the preset mode switching algorithm. The gain parameter is increased for low light environment, and the resolution is reduced for peak period to improve the processing speed, so as to obtain the optimized video processing parameters. The initial video dataset is enhanced using optimized video processing parameters. Noise in low-light environments is removed using an adaptive denoising algorithm, and data redundancy during peak periods is reduced using a fast compression algorithm to obtain the second video dataset. Keyframes were extracted from the second video dataset, and deep learning-based image segmentation techniques were used to separate foreground objects from the background. Detail preservation enhancement was performed on the foreground objects to obtain the third video dataset. A fused video stream is generated based on the third video dataset. Multi-channel video signals are integrated through a spatiotemporal fusion algorithm. Inter-frame interpolation is optimized for peak data streams, and contrast is enhanced for low-light environments to obtain the final video output. Real-time verification is performed on the final video output. If the processing delay exceeds the preset threshold, the resource allocation ratio is adjusted through the feedback mechanism, and the mode switching and enhancement processing are re-executed to obtain the updated video output. The updated video output is used to predict the scene change trend in the next minute using a recurrent neural network. The preset threshold and mode switching frequency are adjusted to obtain adaptive adjustment parameters for the next round of video data stream processing.

2. The method according to claim 1, characterized in that, The acquisition of video data streams of logistics scenarios involves capturing multi-channel video signals through a preset sensor array, and marking timestamps and light intensity for peak-period data streams and low-light environments respectively, to obtain an initial video dataset, including: The initial video dataset is obtained by capturing multi-channel video signals of logistics scenarios using a pre-set sensor array, marking timestamps and light intensity; If the initial video dataset contains peak period data, then dynamic objects are detected using the inter-frame difference method to obtain a sequence of dynamic objects; Based on the dynamic object sequence, the optical flow method is used to calculate the object's motion trajectory, resulting in a set of motion trajectories; If the density of trajectories in the motion trajectory set exceeds a preset threshold, then a clustering algorithm is used to divide the congested areas and obtain the distribution of congested areas. Based on the distribution of congested areas, video frames under low light conditions are obtained, and histogram equalization is used to enhance image contrast, resulting in an enhanced image set. By enhancing the image set, extracting keyframe features, and using a preset template matching method to identify logistics targets, the target recognition results are obtained. Based on the target recognition results, timestamps and light intensity are combined to generate a logistics scenario status dataset.

3. The method according to claim 1, characterized in that, The process involves extracting environmental features from the initial video dataset, using a convolutional neural network to analyze inter-frame illumination changes and object movement speed, determining whether the current scene is a peak or low-light environment, and obtaining environmental change perception parameters, including: Frame sequences were extracted from a video dataset, and the inter-frame difference method was applied to calculate illumination differences and object motion trajectories, thereby obtaining illumination change features and motion speed features. By processing illumination variation features through a convolutional neural network, the spatial distribution pattern of illumination between frames is extracted to obtain illumination distribution parameters; By processing motion velocity features through convolutional neural networks, the temporal variation pattern of object motion is analyzed to obtain motion distribution parameters; If the illumination distribution parameters are lower than the preset threshold, it is judged as a low illumination environment, and low illumination perception parameters are generated. If the motion distribution parameters are higher than the preset threshold, it is determined to be a peak period scene, and peak period perception parameters are generated. Based on low-light perception parameters and peak-hour perception parameters, the environmental perception results are fused to generate comprehensive environmental perception parameters. By integrating environmental perception parameters, updating the scene judgment model, and optimizing the feature extraction of illumination differences and motion speed, adaptive environmental perception parameters are obtained.

4. The method according to claim 1, characterized in that, The resource allocation scheme activates the corresponding video processing mode, adjusts the frame rate and resolution through a preset mode switching algorithm, increases the gain parameter for low-light environments, and decreases the resolution during peak periods to improve processing speed, resulting in optimized video processing parameters, including: The current light intensity data is obtained by using an ambient light sensor, and the type of lighting environment is determined by using a preset threshold to obtain a classification result of low light or normal light. If the classification result is low light, the gain parameter is increased by a preset gain adjustment algorithm to generate the first video processing parameters; Obtain the current system load data, use a preset load threshold to determine whether it is a peak period, and obtain the result of peak or off-peak period judgment; If the judgment result indicates a peak period, the resolution is adjusted using a resolution reduction algorithm to generate the second video processing parameters; Based on the first video processing parameters and the second video processing parameters, a parameter fusion algorithm is used to generate comprehensive video processing parameters; The frame rate adjustment module switches the frame rate of the comprehensive video processing parameters to obtain optimized video parameters.

5. The method according to claim 1, characterized in that, The initial video dataset is enhanced using optimized video processing parameters. Noise in low-light environments is removed using an adaptive denoising algorithm, and data redundancy during peak periods is reduced using a fast compression algorithm, resulting in a second video dataset, which includes: The initial video dataset is processed by an adaptive denoising algorithm. For noise in low-light environments, the weighted average of the gray value of a pixel and its neighboring pixels is calculated. If the gray value deviation exceeds a preset threshold, it is replaced with the weighted average to obtain the first intermediate video dataset after denoising. The first intermediate video dataset is processed using a fast compression algorithm. The pixel change rate between frames is analyzed. If the change rate is lower than a preset threshold, similar frames are merged to reduce data redundancy and obtain the compressed second intermediate video dataset. Based on the inter-frame features of the second intermediate video dataset, edge information of keyframes is extracted, and an edge intensity map is generated through an edge detection algorithm to obtain a third intermediate video dataset containing edge features. For the edge features of the third intermediate video dataset, an adaptive sharpening algorithm is applied to enhance the edge contrast of key frames. If the edge intensity is lower than the preset threshold, the sharpening weight is increased to obtain the sharpened fourth intermediate video dataset. Obtain the brightness distribution of the fourth intermediate video dataset, analyze the pixel values ​​in low-light areas, and if the pixel values ​​are lower than the preset brightness threshold, adjust the brightness through gamma correction to obtain the fifth intermediate video dataset with enhanced brightness. By using the frame sequence of the fifth intermediate video dataset, inter-frame motion vectors are detected, and a motion compensation algorithm is used to optimize the smoothness of the inter-frames. If the motion vector deviation exceeds a preset threshold, a compensation frame is inserted to obtain a smooth second video dataset.

6. The method according to claim 1, characterized in that, The process involves extracting keyframes from the second video dataset, using deep learning-based image segmentation techniques to separate foreground objects and background, and performing detail-preserving enhancement on the foreground objects to obtain the third video dataset, which includes: Frame sequences were obtained from the second video dataset. A pre-trained deep convolutional neural network model was used to extract keyframes through inter-frame difference analysis, resulting in a set of keyframes. A semantic segmentation algorithm is used to process each frame of the keyframe set, and a deep learning model is used to separate the foreground object and the background to generate a foreground mask and a background mask. For the foreground mask region, a super-resolution algorithm is applied to enhance details and generate an enhanced foreground image; If the pixel value of the foreground mask area is lower than the preset threshold, texture restoration is performed on the area to obtain the restored foreground image; By fusing the repaired foreground image with the background mask region, an enhanced complete frame image is generated, resulting in an enhanced frame set; Frame sequences are obtained from the enhanced frame set, reassembled according to timestamp order, and a third video dataset is generated. Based on the inter-frame consistency of the third video dataset, an optical flow algorithm is used to optimize the inter-frame transition, resulting in a smooth third video dataset.

7. The method according to claim 1, characterized in that, The process of generating a fused video stream based on a third video dataset, integrating multi-channel video signals through a spatiotemporal fusion algorithm, optimizing inter-frame interpolation for peak data streams, and enhancing contrast in low-light environments to obtain the final video output includes: Raw data is acquired from multi-channel video signals, and the signals are integrated using a spatiotemporal fusion algorithm to generate a fused video stream; Inter-frame interpolation is used to process the fused video stream, and optimization is performed on the peak data stream to obtain a smooth video sequence; If the smooth video sequence is in a low-light environment, the brightness and contrast are adjusted by a contrast enhancement algorithm to obtain an enhanced video sequence. By conducting environmental adaptability testing, it is determined whether the enhanced video sequence meets the output quality requirements. If not, the contrast parameter is iteratively adjusted to obtain an optimized video sequence. Based on the optimized video sequence, key frame data is obtained, and redundant information is compressed using data stream processing technology to generate a compressed video stream. For compressed video streams, output quality enhancement processing is performed, and noise is removed using a denoising algorithm to obtain the final video output; Feature data is extracted from the final video output and stored in a preset database for subsequent signal integration.

8. The method according to claim 1, characterized in that, The real-time verification of the final video output, if the processing delay exceeds a preset threshold, involves adjusting the resource allocation ratio through a feedback mechanism, re-executing mode switching and enhancement processing, and obtaining an updated video output, including: Real-time video data is obtained from the video stream, and the processing delay is calculated using the inter-frame difference method to obtain the delay value; If the delay value exceeds the preset threshold, a resource adjustment signal is generated through a preset feedback mechanism to determine the resource allocation ratio; Based on the resource adjustment signal, a dynamic scheduling algorithm is used to reallocate computing resources to obtain an updated resource allocation scheme. For the updated resource allocation scheme, perform a mode switching operation to generate the new processing mode; By switching to the new processing mode, an image enhancement algorithm is used to process the video data to obtain enhanced video frames. Extract output data from the enhanced video frames to generate an updated video output; If the updated video output delay still exceeds the preset threshold, return to the resource adjustment signal generation step and repeat until the delay meets the threshold to obtain the final video output.

9. The method according to claim 1, characterized in that, The process involves using a recurrent neural network to predict scene change trends within the next minute through the updated video output, adjusting preset thresholds and mode switching frequencies to obtain adaptive adjustment parameters for the next round of video data stream processing, including: Real-time scene frame sequences are obtained from video stream data, and dynamic features are extracted using the frame difference method to obtain a scene change feature set. For the scene change feature set, a recurrent neural network is used to perform time series analysis to generate the scene change trend for the next minute; If the scene change trend exceeds the preset threshold, the preset threshold is adjusted by gradient descent to obtain the updated threshold parameter. Based on the updated threshold parameters, calculate the mode switching frequency and generate an adaptive switching frequency value; By adaptively switching the frequency value to adjust the video stream processing mode, an optimized processing mode configuration is obtained; The system employs an optimized processing mode to process subsequent video stream data and generate real-time processing results. If the real-time processing results deviate significantly from the changing trend of the scene, the parameters of the recurrent neural network are updated through a feedback mechanism to obtain an improved prediction model.

Images