Improved low-computing-power monitoring target recognition algorithm of YOLO

Abstract

The application discloses an improved low-computing-power monitoring target recognition algorithm of YOLO, and comprises the following steps: S1, replacing the original YOLO main body with a light feature extraction structure, setting the first layer convolution kernel as 16 or 24, and the deep layer convolution kernel being not more than 128, the application has the beneficial effects that: by limiting the upper limit of the convolution kernel quantity and combining the channel pruning based on the scaling factor of the normalization layer, the low-contribution connection in the network is removed, the compression of the inference engine volume is realized; by utilizing the characteristics that the interframe correlation of the monitoring video is high, an interframe feature cache area is constructed, only full feature extraction is performed on the first frame, the deep layer features are reused and only the shallow layer convolution is performed on the subsequent frames, and the repeated calculation is reduced; by means of a non-uniform precision distribution strategy, high-precision floating point operation is adopted in the key area, low-bit integer operation is adopted in the non-key area, and memory alignment management is matched, so that the operation cost is reduced from three aspects of structure, scheduling and value.

Description

Technical Field

[0001] This invention relates to the field of target recognition technology, specifically to an improved YOLO low-computing-power monitoring target recognition algorithm. Background Technology

[0002] With the advancement of smart city and security monitoring construction, deep learning-based object detection algorithms have been widely applied in the field of video surveillance; YOLO (You Only Look) YOLO series algorithms have become the mainstream detection framework due to their fast detection speed and high accuracy. However, existing YOLO algorithms typically involve a large number of parameters and computational loads, placing high demands on the computing power and storage resources of the deployed hardware. Although deploying the algorithm on cloud servers can solve the problem of insufficient computing power at a single point, the network bandwidth pressure and data transmission latency brought about by the massive uploading of surveillance videos make edge inference an inevitable trend in real-time monitoring. In order to run target detection algorithms on embedded edge devices (such as ARM architecture chips, edge computing boxes, etc.), the industry usually adopts model pruning, quantization, and lightweight network design. However, current pruning methods often lack accurate assessment of the importance of channels, and blindly removing channels can easily lead to a decline in network representation capabilities. Low-bit quantization technology lacks simulation of the operational characteristics of low-computing-power chips before the deployment of the inference engine. Direct quantization often causes numerical overflow or a sharp drop in accuracy, making it difficult to maintain recognition performance at extremely low bit depths. Existing detection methods usually perform full feature extraction and calculation for each frame of image, failing to effectively utilize the redundancy of inter-frame information. This results in a large amount of computing power being consumed in the repeated analysis of static backgrounds, which greatly limits the effective processing frame rate of low-computing-power chips. Summary of the Invention

[0003] The purpose of this invention is to provide an improved YOLO low-computing-power target recognition algorithm to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an improved YOLO low-computing-power monitoring target recognition algorithm, comprising the following steps: S1. Replace the original YOLO backbone with a lightweight feature extraction structure, set the first layer convolutional kernel to 16 or 24, and the deep convolutional kernel to no more than 128; remove cross-stage physical connection paths and adopt depthwise separable convolution. S2. Add a region attention masking module to the output of S1; generate a static mask matrix using the previous few frames, align it by bilinear interpolation, and assign a weight of 0.01-0.05 to the low-frequency region and a weight of 1.0 to the high-frequency region. S3. Quantize the sum of absolute values ​​of pixel differences between adjacent frames to determine the intensity of motion, and adjust the resolution adaptively according to the lighting. Extract and cache deep features in the first frame, refresh every 15-30 frames or when the difference exceeds the 5‰ threshold, and perform shallow convolution only in subsequent frames. S4. Based on the feature reuse of S3, short-term motion trend is introduced to assist detection; the target position and displacement vector of the previous frame are recorded, and the candidate box of the current frame is calculated assuming uniform speed. S5. Construct a resource monitoring thread to collect processor utilization, memory read / write bandwidth, and graphics computing unit queue length, and obtain a comprehensive computing power score by weighting the data at 0.4, 0.3, and 0.3. Based on this, switch the post-processing strategy, combine the scale compensation coefficient to correct the detection deviation, and merge the bounding box decoding, scale compensation, and filtering into a single graphics computing unit kernel startup. S6. Before deployment, simulate a low-bit computing environment, apply quantization noise to the S1 structure and retrain; combine with the S2 static mask, use floating point in the high-weight area and low-bit integer in the low-weight area to perform non-uniform precision allocation and isolate memory; fine-tune the quantization parameters within ±5% based on the S5 score to prevent overflow.

[0005] Preferably, step S1 specifically includes the following steps: S11. Replace the backbone of the original YOLO detection architecture with a feature extraction structure with fewer parameters. The feature extraction structure is composed of multiple lightweight convolutional units connected in series. The number of convolutional kernels in the first layer is set to 16 or 24, and the number of convolutional kernels in the deep layer does not exceed 128. S12. Remove the cross-stage physical connection path with high computational overhead in the feature extraction structure and replace it with a direct residual connection. S13. Replace the original standard convolution operation in the feature extraction structure with depthwise separable convolution operation, and decouple the originally coupled spatial feature extraction and channel feature fusion into two independent operation steps. S14. After training converges, the importance of channels is evaluated based on the scaling factor of the normalization layer output, and redundant channels with scaling factors close to zero are removed. S15. Fine-tune the weight parameters of the remaining channels with a small number of iterations. The fine-tuning stage uses a stochastic gradient descent optimizer, with the learning rate set to less than one-tenth of that in the initial training stage, and the batch size set to 32 or 64.

[0006] Preferably, step S2 specifically includes the following steps: S21. A region attention masking module is added to the output end of the feature extraction structure described in S1; for the frame sequence acquired by the fixed acquisition device, the distribution frequency of the foreground pixels in the first few frames is accumulated to construct a heat map reflecting the probability of the target's appearance; the heat map is binarized to generate a static mask matrix, wherein the binarization threshold is set to 20% to 30% of the maximum probability value of the heat map; S22. Since the spatial resolution of the static mask matrix is ​​usually lower than the multi-channel feature map output by the feature extraction structure, bilinear interpolation is first used to upsample the static mask matrix and scale it to the same spatial resolution and size as the feature map. S23. Perform element-wise multiplication of the aligned mask values ​​with all channel values ​​at the corresponding spatial positions on the feature map; assign values ​​differently according to the importance of the regions. For the low-frequency regions identified by the mask, set the mask values ​​corresponding to the low-frequency regions to 0.01 to 0.05; for the high-frequency regions identified by the mask, set the mask values ​​corresponding to the high-frequency regions to 1.0.

[0007] Preferably, step S3 specifically includes the following steps: S31. The intensity of motion in the monitored image is quantified by calculating the sum of the absolute values ​​of the pixel differences between two adjacent frames. The preset threshold used to determine the motion state is adaptively determined based on the average light intensity of the monitored scene. The higher the light intensity, the higher the preset threshold is. S32. Dynamically adjust the input resolution based on the quantization result of S31; when the inter-frame difference is lower than the preset threshold, it is determined to be a static state and the input pixels are automatically reduced; when the inter-frame difference is higher than the preset threshold, it is determined to be a violent motion state and the high-resolution input is restored. S33. Construct an inter-frame feature buffer. For a continuously input frame sequence, only the first frame in the sequence is called to perform full feature extraction from shallow to deep layers using the feature extraction structure described in S1, and the resulting deep semantic feature map is stored in the inter-frame feature buffer. S34. Set cache validity rules, and the inter-frame feature cache area is forcibly refreshed every 15 to 30 frames; at the same time, when the difference value between consecutive frames exceeds the high resolution switching threshold, the cache refresh is immediately triggered. The high resolution switching threshold is set to the sum of the absolute values ​​of the pixel differences in a single frame exceeding 5‰ of the total number of pixels in the whole frame. S35. When processing subsequent frames, the full feature extraction is no longer repeated. Instead, the deep semantic feature map in the inter-frame feature buffer is read directly, and shallow convolution is performed only on the newly input current frame to extract edge and texture features.

[0008] Preferably, step S4 specifically includes the following steps: S41. Based on the feature reuse mechanism described in S3, an auxiliary detection mechanism based on short-term motion trends is introduced; for a continuous frame sequence reused in S3, a target tracking list is continuously maintained, specifically recording the position coordinates and displacement vectors of the target that has been successfully locked in the previous frame. S42. Assuming that the target in the monitored scene maintains uniform linear motion for a short period of time, based on the displacement vector recorded in S41, the expected position of the target in the current frame is calculated; the target center coordinates and displacement vector of the previous frame are added together to obtain the predicted center coordinates of the current frame, and a candidate region suggestion box is generated with the predicted center coordinates as the center. S43. When the target response weakens in the current frame due to the reduction of input resolution or extraction of only shallow features in S3, a feature enhancement strategy is initiated; the spatial position of the candidate region proposal box generated in S42 on the feature map is located, and the feature map amplitude within the candidate region is multiplied by an enhancement coefficient in the range of 1.5 to 3.0. S44. The steps described above work in conjunction with the dynamic reasoning mechanism of S3. Feature enhancement is triggered when S3 determines that the calculation is in a static or shallow calculation mode, and automatically goes into hibernation during high-resolution full calculation.

[0009] Preferably, step S5 specifically includes the following steps: S51. Construct a resource monitoring thread for the edge device. The resource monitoring thread runs independently of the inference main process and periodically collects the real-time operating status of the chip. The collected indicators specifically include processor utilization, memory read / write bandwidth, and task queue length of the graphics computing unit. The three indicators are weighted and summed to generate a comprehensive computing power score. The weighting coefficients for processor utilization, memory read / write bandwidth, and task queue length of the graphics computing unit are set to 0.4, 0.3, and 0.3, respectively. S52. Based on the comprehensive computing power score generated in S51, dynamically adjust the execution strategy of the post-processing pipeline; when the comprehensive computing power score is lower than the preset threshold, enable the lightweight mode, simplify the candidate box screening process based on intersection-union ratio to directly extract the top K candidate boxes after sorting by confidence score, and temporarily lower the confidence threshold; when the comprehensive computing power score is higher than the preset threshold, resume frame-by-frame intersection-union comparison screening, and adopt a strict confidence threshold. S53. To address the candidate box scale differences caused by dynamic resolution switching in S3, a scale compensation coefficient is introduced in the post-processing stage. The corresponding scaling factor needs to be calculated based on the ratio between the current input resolution and the original input image resolution. Using the scaling factor, the coordinates of the candidate boxes generated at different resolutions are mapped back to the scale space of the original input image. S54. Refactor the post-processing code logic, merging bounding box decoding, scale compensation, and candidate box filtering into a kernel startup process for a single graphics calculation unit.

[0010] Preferably, step S6 specifically includes the following steps: S61. Before deploying the inference engine, perform a mixed precision conversion process. For the feature extraction structure after pruning in S1, simulate the low-bit integer operation characteristics of low-computing-power chips during the forward inference process. Before the weight parameters and activation values ​​enter the convolution kernel, artificially apply quantization noise. Then, retrain the full dataset. S62. Combine the static mask matrix generated in S2 and implement a non-uniform precision allocation strategy; count the importance of each position along the spatial dimension of the feature map, define the region marked with high weight by the static mask as the high-precision calculation region, and define the remaining region as the low-precision calculation region. S63. For the different precision regions defined above, implement differentiated memory management during the inference initialization phase; specify the input and output tensors of the computation layer corresponding to the high-precision computation region as higher-precision floating-point format; specify the tensors corresponding to the low-precision computation region as low-bit integer format; and allocate independent memory alignment blocks for these two types of tensors of different data types respectively. S64. Establish a closed-loop adjustment mechanism for quantization parameters and hardware load; based on the comprehensive computing power score generated in S5, dynamically fine-tune the quantization cutoff threshold and scaling factor of each layer; set constraints to limit the fine-tuning range to ±5% of the initial quantization parameters. S65, the steps mentioned above form a close upstream and downstream linkage with the structural compression of S1 and the hardware-aware scheduling of S5; S1 reduces the total amount of weight parameters that need to be quantized through pruning.

[0011] Preferably, the specific operations for evaluating the importance of the channel in step S14 include: Using the scaling factor output by the normalization layer as the channel importance weight coefficient, calculate the L1 norm or mean of the scaling factor; set the channel pruning rate, sort the scaling factor values ​​in ascending order, and select the channel with the smallest value as a redundant channel for removal; the removal operation includes setting the convolution kernel weight parameter of the redundant channel in the feature extraction structure to zero, and updating the channel index mapping table of subsequent calculation layers.

[0012] Preferably, in step S34, the triggering logic for the cache validity rule specifically includes: A frame counter and a differential accumulator are set. The frame counter increments by 1 after each frame is processed. When the frame counter value reaches the upper limit of the range of 15 to 30 frames, a forced refresh signal is triggered. The differential accumulator accumulates the sum of the absolute values ​​of the pixel differences between adjacent frames. When the differential accumulator value exceeds the high resolution switching threshold, an instant refresh signal is triggered. After any refresh signal is triggered, the frame counter and the differential accumulator are reset, and the feature extraction structure described in S1 is called to re-execute the full feature extraction.

[0013] Preferably, the specific implementation of the single graphics computing unit kernel startup process in step S54 includes: In the kernel startup function, a shared memory buffer is defined, and the original coordinate data generated by the bounding box decoding, the scaling parameters calculated by the scale compensation coefficient, and the confidence threshold for candidate box filtering are passed in as input parameters all at once. Inside the graphics computing unit, coordinate inverse normalization operation, scale space mapping operation, and confidence-based candidate box filtering operation are executed sequentially. The operation results are directly written to the output video memory, and intermediate results are not copied or synchronized between the processor and the graphics computing unit.

[0014] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: By limiting the upper limit of the number of convolutional kernels and combining channel pruning based on the normalization layer scaling factor, low-contribution connections in the network are removed, thereby compressing the size of the inference engine; taking advantage of the high inter-frame correlation of surveillance video, an inter-frame feature buffer is constructed, and full feature extraction is performed only in the first frame, while deep features are reused and only shallow convolution is performed in subsequent frames, reducing redundant computation; through a non-uniform precision allocation strategy, high-precision floating-point operations are used in critical areas, and low-bit integer operations are used in non-critical areas, combined with memory alignment management, reducing computational overhead from three levels: structure, scheduling, and numerical values, enabling the algorithm to run on embedded devices with limited storage and computing power; By constructing a hardware resource monitoring thread independent of the inference main process, processor utilization, memory bandwidth, and graphics computing unit queue length are collected in real time, and a comprehensive computing power score is generated based on the weighted average. The post-processing strategy is dynamically switched accordingly. Under high load, the filtering logic is simplified to maintain the frame rate, and under low load, fine filtering is restored to ensure accuracy. Bounding box decoding, scale compensation, and candidate box filtering are merged into a single graphics computing unit kernel startup, and parameters are passed in at once using a shared memory buffer, avoiding frequent intermediate result copying and synchronization waiting between the processor and the graphics computing unit. Attached Figure Description

[0016] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0018] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses consistent with some aspects of this disclosure as detailed in the appended claims.

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Example 1

[0020] See Figure 1 As shown, an improved YOLO low-computing-power monitoring target recognition algorithm according to an embodiment of the present invention includes the following steps: S1. Replace the original YOLO backbone with a lightweight feature extraction structure, set the first layer convolutional kernel to 16 or 24, and the deep convolutional kernel to no more than 128; remove cross-stage physical connection paths and adopt depthwise separable convolution. S2. Add a region attention masking module to the output of S1; use the previous few frames to generate a static mask matrix, and after bilinear interpolation alignment, assign a weight of 0.01-0.05 to the low frequency region and a weight of 1.0 to the high frequency region. S3. Quantize the sum of absolute values ​​of pixel differences between adjacent frames to determine the intensity of motion, and adjust the resolution adaptively according to the lighting. Extract and cache deep features in the first frame, refresh every 15-30 frames or when the difference exceeds the 5‰ threshold, and perform shallow convolution only in subsequent frames. S4. Based on the feature reuse of S3, short-term motion trend is introduced to assist detection; the target position and displacement vector of the previous frame are recorded, and the candidate box of the current frame is calculated assuming uniform speed. S5. Construct a resource monitoring thread to collect processor utilization, memory read / write bandwidth, and graphics computing unit queue length, and obtain a comprehensive computing power score by weighting the data at 0.4, 0.3, and 0.3. Based on this, switch the post-processing strategy, combine the scale compensation coefficient to correct the detection deviation, and merge the bounding box decoding, scale compensation, and filtering into a single graphics computing unit kernel startup. S6. Before deployment, simulate a low-bit computing environment, apply quantization noise to the S1 structure and retrain; combine with the S2 static mask, use floating point in the high-weight area and low-bit integer in the low-weight area to perform non-uniform precision allocation and isolate memory; fine-tune the quantization parameters within ±5% based on the S5 score to prevent overflow. Example 2

[0021] Step S1 specifically includes the following steps: S11. Replace the backbone of the original YOLO detection architecture with a feature extraction structure with fewer parameters. The feature extraction structure is composed of multiple lightweight convolutional units connected in series. The number of convolutional kernels in the first layer is set to 16 or 24, and the number of convolutional kernels in the deep layer does not exceed 128. S12. Remove the computationally expensive cross-stage physical connection paths in the feature extraction structure and replace them with direct residual connections. S13. Replace the original standard convolution operation in the feature extraction structure with depthwise separable convolution operation, and decouple the originally coupled spatial feature extraction and channel feature fusion into two independent operation steps. S14. After training converges, the importance of channels is evaluated based on the scaling factor of the normalization layer output, and redundant channels with scaling factors close to zero are removed. S15. Fine-tune the weight parameters of the remaining channels in a small number of iterations. The fine-tuning stage uses a stochastic gradient descent optimizer, with the learning rate set to less than one-tenth of that in the initial training stage, and the batch size set to 32 or 64. By limiting the upper limit of the number of convolutional kernels in the first and deep layers, the number of model parameters and computational complexity are restricted from the source. By limiting the learning rate and batch size in the fine-tuning stage, the distribution drift of weight parameters after pruning is suppressed, solving the problem of the model being difficult to converge or severely impaired in accuracy after deep pruning. This structured compression method provides a numerically stable foundation for subsequent low-bit quantization, enabling the inference engine to maintain usable recognition performance under extremely compressed conditions, and adapting to the storage and computing power limits of embedded devices.

[0022] Step S2 specifically includes the following steps: S21. Add a region attention masking module to the output of the feature extraction structure in S1; for the frame sequence acquired by the fixed acquisition device, accumulate the distribution frequency of the foreground pixels in the first few frames to construct a heat map reflecting the probability of the target's appearance; binarize the heat map to generate a static mask matrix, wherein the binarization threshold is set to 20% to 30% of the maximum probability value of the heat map; S22. Since the spatial resolution of the static mask matrix is ​​usually lower than that of the multi-channel feature map output by the feature extraction structure, bilinear interpolation is first used to upsample the static mask matrix and scale it to the same spatial resolution and size as the feature map. S23. Perform element-wise multiplication of the aligned mask values ​​with all channel values ​​at the corresponding spatial positions on the feature map; assign values ​​differently according to the importance of the regions. For low-frequency regions identified by the mask, set the corresponding mask values ​​for the low-frequency regions to 0.01 to 0.05; for high-frequency regions identified by the mask, set the corresponding mask values ​​for the high-frequency regions to 1.0. Leveraging prior knowledge that the background in surveillance video is relatively static, a static mask matrix is ​​constructed by statistically analyzing the first few frames, and bilinear interpolation is used to achieve geometric alignment between the mask and the feature map. Without introducing additional trainable parameters, attention modulation of the spatial dimension is achieved through channel-wise product operations. By setting mask weights for low-frequency and high-frequency regions differently, the response of invalid regions such as the sky and building exteriors can be effectively suppressed, guiding computing resources to tilt towards key areas such as roads and entrances / exits, thereby improving the efficiency of computing resource allocation.

[0023] Step S3 specifically includes the following steps: S31. The intensity of motion in the monitored image is quantified by calculating the sum of the absolute values ​​of the pixel differences between two adjacent frames. The preset threshold used to determine the motion state is adaptively determined based on the average light intensity of the monitored scene. The higher the light intensity, the higher the preset threshold. S32. Dynamically adjust the input resolution based on the quantization result of S31; when the inter-frame difference is lower than the preset threshold, it is determined to be a static state and the input pixels are automatically reduced; when the inter-frame difference is higher than the preset threshold, it is determined to be a violent motion state and the high-resolution input is restored. S33. Construct an inter-frame feature buffer. For a continuously input frame sequence, only call the feature extraction structure of S1 to perform full feature extraction from shallow to deep layers on the first frame of the sequence, and store the obtained deep semantic feature map into the inter-frame feature buffer. S34. Set cache validity rules, and force refresh the inter-frame feature cache every 15 to 30 frames; at the same time, when the difference between consecutive frames exceeds the high resolution switching threshold, the cache refresh is triggered immediately. The high resolution switching threshold is set to the sum of the absolute values ​​of the pixel differences in a single frame exceeding 5‰ of the total number of pixels in the whole frame. S35. When processing subsequent frames, the full feature extraction is no longer repeated. Instead, the deep semantic feature map in the inter-frame feature buffer is read directly, and shallow convolution is performed only on the current frame of the new input to extract edge and texture features. By introducing a dynamic threshold adapted to illumination and a dual refresh mechanism, the management strategy for inter-frame feature caching is optimized. By setting the collaborative logic between the frame counter and the differential accumulator, it is possible to track slow changes in the environment and respond promptly to sudden target movements. The correlation between surveillance video frames is fully utilized, avoiding repeated extraction of deep semantic features during periods of background stillness and reducing the average computational load.

[0024] Step S4 specifically includes the following steps: S41. Based on the feature reuse mechanism in S3, an auxiliary detection mechanism based on short-term motion trends is introduced; for a continuous frame sequence reused in S3, a target tracking list is continuously maintained, specifically recording the position coordinates and displacement vectors of the target that has been successfully locked in the previous frame. S42. Assuming that the target in the monitored scene maintains uniform linear motion for a short period of time, based on the displacement vector recorded in S41, calculate the expected position of the target in the current frame; add the target center coordinates and displacement vector of the previous frame to obtain the predicted center coordinates of the current frame, and generate candidate region suggestion boxes with the predicted center coordinates as the center. S43. When the target response weakens in the current frame due to the reduced input resolution in S3 or the extraction of only shallow features, the feature enhancement strategy is activated; the spatial position of the candidate region proposal box generated in S42 on the feature map is located, and the feature map amplitude within the candidate region is multiplied by an enhancement coefficient in the range of 1.5 to 3.0. S44, the steps work in conjunction with the dynamic reasoning mechanism of S3. Feature enhancement is triggered when S3 determines that it is in a static or shallow calculation mode, and automatically goes into hibernation during high-resolution full calculation. Compared to computationally intensive optical flow algorithms, this method can predict target positions using only coordinate calculations, reducing the computational overhead of temporal correlation. By applying feature enhancement coefficients to candidate regions in low-resolution or shallow computation modes, it can compensate for the feature response decay caused by reduced information, which helps maintain the continuity of target locking and reduces the risk of missed detections during dynamic resolution switching.

[0025] Step S5 specifically includes the following steps: S51. Construct a resource monitoring thread for the edge device. The resource monitoring thread runs independently of the inference main process and periodically collects the real-time operating status of the chip. The specific indicators collected include processor utilization, memory read / write bandwidth, and task queue length of the graphics computing unit. The three indicators are weighted and summed to generate a comprehensive computing power score. The weighting coefficients for processor utilization, memory read / write bandwidth, and task queue length of the graphics computing unit are set to 0.4, 0.3, and 0.3, respectively. S52. Based on the comprehensive computing power score generated by S51, dynamically adjust the execution strategy of the post-processing pipeline; when the comprehensive computing power score is lower than the preset threshold, activate the lightweight mode, simplify the candidate box screening process based on intersection-union ratio to directly extract the top K candidate boxes after sorting by confidence score, and temporarily lower the confidence threshold; when the comprehensive computing power score is higher than the preset threshold, restore the frame-by-frame intersection-union comparison screening, and adopt a strict confidence threshold. S53. To address the candidate box scale differences caused by dynamic resolution switching in S3, a scale compensation coefficient is introduced in the post-processing stage. The corresponding scaling factor needs to be calculated based on the ratio between the current input resolution and the original input image resolution. Using the scaling factor, the coordinates of the candidate boxes generated at different resolutions are mapped back to the scale space of the original input image. S54. Refactor the post-processing code logic, merging bounding box decoding, scale compensation, and candidate box filtering into a kernel startup process for a single graphics calculation unit; A multi-dimensional hardware load awareness model was constructed, which collects three heterogeneous indicators: processor utilization, memory bandwidth, and graphics computing unit queue length, and generates a comprehensive computing power score according to specific weight coefficients. The score reflects the chip's true load status better than a single indicator. Based on this score, the post-processing strategy is dynamically switched, which can flexibly balance detection accuracy and processing speed: the frame rate is maintained by simplifying the screening logic under high load, and the accuracy is guaranteed by fine comparison under low load, thus realizing the elastic scheduling of computing resources under different working conditions.

[0026] Step S6 specifically includes the following steps: S61. Before deploying the inference engine, perform a mixed precision conversion process. For the feature extraction structure after pruning in S1, simulate the low-bit integer operation characteristics of low-computing-power chips during the forward inference process. Before the weight parameters and activation values ​​enter the convolution kernel, artificially apply quantization noise. Then, retrain the full dataset. S62. Combine the static mask matrix generated in S2 to implement a non-uniform precision allocation strategy; count the importance of each position along the spatial dimension of the feature map, define the region marked with high weight by the static mask as the high-precision calculation region, and define the remaining region as the low-precision calculation region. S63. For the different precision regions defined above, implement differentiated memory management during the inference initialization phase; specify the input and output tensors of the computation layer corresponding to the high-precision computation region as higher-precision floating-point format; specify the tensors corresponding to the low-precision computation region as low-bit integer format; and allocate independent memory alignment blocks for these two types of tensors of different data types respectively. S64. Establish a closed-loop adjustment mechanism for quantization parameters and hardware load; based on the comprehensive computing power score generated in S5, dynamically fine-tune the quantization cutoff threshold and scaling factor of each layer; set constraints to limit the fine-tuning range to ±5% of the initial quantization parameters. S65, the structural compression of steps and S1, and the hardware-aware scheduling of S5 form a close upstream and downstream linkage; S1 reduces the total number of weight parameters that need to be quantized through pruning. By simulating low-bit operation characteristics and applying quantization noise in forward inference, a quantization-aware training process was constructed to alleviate the accuracy loss caused by low-bit quantization. Combined with a non-uniform precision allocation strategy implemented by static masking, higher precision is retained in critical areas while low-bit operations are used in non-critical areas, achieving a fine trade-off between computational accuracy and throughput. A dynamic fine-tuning mechanism based on real-time computing power scoring enables the quantization parameters to be adaptively corrected according to the actual hardware load.

[0027] In step S14, the specific operations for assessing the importance of a channel include: The scaling factor output by the normalization layer is used as the channel importance weight coefficient. The L1 norm or mean of the scaling factor is calculated. The channel pruning rate is set, the scaling factor values ​​are sorted in ascending order, and the channel with the smallest value is selected as the redundant channel for removal. The removal operation includes setting the convolution kernel weight parameter of the redundant channel in the feature extraction structure to zero and updating the channel index mapping table of the subsequent calculation layer. By using the scaling factor as the criterion for channel weights and sorting and filtering using the L1 norm or mean, the sparsity of the model is automated and quantified. After removing redundant channels, the logical correctness of the network topology is ensured by updating the channel index mapping table of subsequent computation layers. This avoids the blindness of manually setting pruning rules and provides a feasible technical path for generating compact and efficient models.

[0028] Specifically, in step S34, the triggering logic for the cache validity rule includes: A frame counter and a differential accumulator are set. The frame counter increments by 1 after each frame is processed. When the frame counter value reaches the upper limit of the range of 15 to 30 frames, a forced refresh signal is triggered. The differential accumulator accumulates the sum of the absolute values ​​of the pixel differences between adjacent frames. When the differential accumulator value exceeds the high resolution switching threshold, an instant refresh signal is triggered. After any refresh signal is triggered, the frame counter and the differential accumulator are reset, and the feature extraction structure of S1 is called to re-execute the full feature extraction. The cache refresh strategy is optimized through a dual triggering mechanism of frame counter and differential accumulator. The forced refresh mechanism can cope with slowly changing environmental factors such as gradual changes in lighting, while the instant refresh mechanism can respond to sudden movements. This combined strategy minimizes the number of full feature extractions while effectively preventing detection failures caused by outdated cached data.

[0029] Specifically, step S54, the implementation of the single graphics computing unit kernel startup process includes: In the kernel startup function, a shared memory buffer is defined, and the original coordinate data generated by the bounding box decoding, the scaling parameters for the scale compensation coefficient calculation, and the confidence threshold for candidate box filtering are passed in as input parameters all at once. Inside the graphics computing unit, coordinate inverse normalization operation, scale space mapping operation, and confidence-based candidate box filtering operation are executed in sequence. The operation results are directly written to the output video memory, and intermediate results are not copied or synchronized between the processor and the graphics computing unit. By refactoring the post-processing code logic, bounding box decoding, scale compensation, and candidate box filtering are merged into a single graphics computing unit kernel startup; by defining a shared memory buffer internally and passing in the relevant parameters all at once, frequent intermediate result copying and synchronization waiting between the processor and the graphics computing unit are avoided. Example 3

[0030] During the model building phase, the algorithm performs lightweight reconstruction of the backbone of the original detection architecture. By reducing the size of the convolutional kernel and removing high-overhead cross-stage physical connections, it replaces them with direct residual connections, thereby reducing the number of parameters and data transfer costs from the source. The scaling factor of the normalization layer output is used as the channel importance criterion to remove redundant channels with values ​​close to zero, and the retained weight parameters are fine-tuned with a small learning rate. This process achieves significant compression of the inference engine size without sacrificing the original recognition accuracy, making it adaptable to the storage and computing power limits of embedded devices. In response to the characteristic that the background occupies most of the pixels in the monitoring image, a spatial attention mechanism based on scene prior is introduced.A static mask matrix is ​​generated by statistically analyzing the foreground distribution of historical frames, and bilinear interpolation is used to achieve geometric alignment with the feature map. Low-frequency regions defined by the mask are assigned extremely low weights, while high-frequency regions are assigned constant weights. Channel-wise multiplication operations suppress invalid background responses. This process does not introduce additional trainable parameters but effectively directs computational resources towards key areas such as road surfaces and entrances / exits. Considering the high similarity between surveillance video frames, an inter-frame feature caching mechanism is constructed. Full feature extraction is performed only on the first frame of the sequence, and deep semantic features are stored in the cache. For subsequent frames, the intensity of motion is quantified by calculating the sum of the absolute values ​​of pixel differences, and the input resolution is adaptively adjusted based on lighting conditions. If the image is static, the resolution is reduced and only shallow convolutions are performed to extract edge textures. If the image is in rapid motion, high resolution is restored and a cache refresh is triggered. By setting a dual triggering logic of frame rate period and differential threshold, the timeliness of cached data is ensured. This layered update strategy avoids repeated calculations of background areas and improves the effective processing frame rate. To address the issue of weakened target response caused by low resolution or shallow feature extraction, the algorithm introduces a short-term motion trend assistance mechanism. It records the position and displacement vector of the target in the previous frame, assuming that the target is moving at a constant speed in a straight line for a short period of time. Based on this, the candidate region of the current frame is calculated. When a weakened target response is detected, a gain coefficient is applied to the feature map amplitude within the candidate region. The algorithm enhances the continuity of target locking without introducing complex optical flow calculations, effectively reducing the risk of missed detections during dynamic switching. At the hardware scheduling level, the algorithm constructs a hardware resource monitoring thread independent of the main inference process. This thread collects processor utilization, memory bandwidth, and graphics computing unit queue length in real time, and generates a weighted comprehensive computing power score. Based on this score, the system's dynamic switching post-processing strategy is as follows: simplifying the filtering logic to maintain frame rate under high load, and performing fine-grained comparisons to ensure accuracy under low load. Simultaneously, to address scale differences caused by dynamic resolution, a scale compensation coefficient is introduced for coordinate mapping. By reconstructing the post-processing code logic, bounding box decoding, scale compensation, and filtering are merged into a single... The secondary graphics computing unit (GPU) kernel startup eliminates frequent data copy latency between the processor and the GPU. During deployment, the algorithm simulates a low-bit computing environment, applies quantization noise to the pruned model, and retrains it to enhance the model's adaptability to low-bit quantization. Combined with static mask information, non-uniform precision allocation is implemented: floating-point format is used in critical areas to maintain precision, while low-bit integer format is used in non-critical areas to improve throughput. Independent memory alignment blocks are allocated for tensors of different precisions. Based on real-time computing power scoring, quantization parameters are dynamically fine-tuned within a limited range to prevent numerical overflow. This further unleashes the hardware potential at the numerical representation level, ensuring recognition stability under extreme low-power modes.

[0031] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0032] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope of this disclosure is indicated by the following claims.

Claims

1. An improved YOLO low-computing-power target recognition algorithm, characterized in that, Includes the following steps: S1. Replace the original YOLO backbone with a lightweight feature extraction structure, set the first layer convolutional kernel to 16 or 24, and the deep layer convolutional kernel to no more than 128. Remove cross-stage physical connection pathways and employ depthwise separable convolution; S2. Add a region attention masking module to the output of S1; use the previous few frames to generate a static mask matrix, and after bilinear interpolation alignment, assign a weight of 0.01-0.05 to the low frequency region and a weight of 1.0 to the high frequency region. S3. Quantize the sum of absolute values ​​of pixel differences between adjacent frames to determine the intensity of motion, and adjust the resolution adaptively according to the lighting. Extract and cache deep features in the first frame, refresh every 15-30 frames or when the difference exceeds the 5‰ threshold, and perform shallow convolution only in subsequent frames. S4. Based on the feature reuse of S3, short-term motion trend is introduced to assist detection; the target position and displacement vector of the previous frame are recorded, and the candidate box of the current frame is calculated assuming uniform speed. S5. Construct a resource monitoring thread to collect processor utilization, memory read / write bandwidth, and graphics computing unit queue length, and obtain a comprehensive computing power score by weighting the data at 0.4, 0.3, and 0.

3. Based on this, switch the post-processing strategy, combine the scale compensation coefficient to correct the detection deviation, and merge the bounding box decoding, scale compensation, and filtering into a single graphics computing unit kernel startup. S6. Before deployment, simulate a low-bit computing environment, apply quantization noise to the S1 structure and retrain; combine with the S2 static mask, use floating point in the high-weight area and low-bit integer in the low-weight area to perform non-uniform precision allocation and isolate memory; fine-tune the quantization parameters within ±5% based on the S5 score to prevent overflow.

2. The improved YOLO low-computing-power monitoring target recognition algorithm according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11. Replace the backbone of the original YOLO detection architecture with a feature extraction structure with fewer parameters. The feature extraction structure is composed of multiple lightweight convolutional units connected in series. The number of convolutional kernels in the first layer is set to 16 or 24, and the number of convolutional kernels in the deep layer does not exceed 128. S12. Remove the cross-stage physical connection path with high computational overhead in the feature extraction structure and replace it with a direct residual connection. S13. Replace the original standard convolution operation in the feature extraction structure with depthwise separable convolution operation, and decouple the originally coupled spatial feature extraction and channel feature fusion into two independent operation steps. S14. After training converges, the importance of channels is evaluated based on the scaling factor of the normalization layer output, and redundant channels with scaling factors close to zero are removed. S15. Fine-tune the weight parameters of the remaining channels with a small number of iterations. The fine-tuning stage uses a stochastic gradient descent optimizer, with the learning rate set to less than one-tenth of that in the initial training stage, and the batch size set to 32 or 64.

3. The improved YOLO low-computing-power monitoring target recognition algorithm according to claim 2, characterized in that, Step S2 specifically includes the following steps: S21. A region attention masking module is added to the output end of the feature extraction structure described in S1; for the frame sequence acquired by the fixed acquisition device, the distribution frequency of the foreground pixels in the first few frames is accumulated to construct a heat map reflecting the probability of the target's appearance; the heat map is binarized to generate a static mask matrix, wherein the binarization threshold is set to 20% to 30% of the maximum probability value of the heat map; S22. Since the spatial resolution of the static mask matrix is ​​usually lower than the multi-channel feature map output by the feature extraction structure, bilinear interpolation is first used to upsample the static mask matrix and scale it to the same spatial resolution and size as the feature map. S23. Perform element-wise multiplication of the aligned mask values ​​with all channel values ​​at the corresponding spatial positions on the feature map; assign values ​​differently according to the importance of the regions. For the low-frequency regions identified by the mask, set the mask values ​​corresponding to the low-frequency regions to 0.01 to 0.05; for the high-frequency regions identified by the mask, set the mask values ​​corresponding to the high-frequency regions to 1.

0.

4. The improved YOLO low-computing-power monitoring target recognition algorithm according to claim 3, characterized in that, Step S3 specifically includes the following steps: S31. The intensity of motion in the monitored image is quantified by calculating the sum of the absolute values ​​of the pixel differences between two adjacent frames. The preset threshold used to determine the motion state is adaptively determined based on the average light intensity of the monitored scene. The higher the light intensity, the higher the preset threshold is. S32. Dynamically adjust the input resolution based on the quantization result of S31; when the inter-frame difference is lower than the preset threshold, it is determined to be a static state and the input pixels are automatically reduced; when the inter-frame difference is higher than the preset threshold, it is determined to be a violent motion state and the high-resolution input is restored. S33. Construct an inter-frame feature buffer. For a continuously input frame sequence, only the first frame in the sequence is called to perform full feature extraction from shallow to deep layers using the feature extraction structure described in S1, and the resulting deep semantic feature map is stored in the inter-frame feature buffer. S34. Set cache validity rules, and the inter-frame feature cache area is forcibly refreshed every 15 to 30 frames; at the same time, when the difference value between consecutive frames exceeds the high resolution switching threshold, the cache refresh is immediately triggered. The high resolution switching threshold is set to the sum of the absolute values ​​of the pixel differences in a single frame exceeding 5‰ of the total number of pixels in the whole frame. S35. When processing subsequent frames, the full feature extraction is no longer repeated. Instead, the deep semantic feature map in the inter-frame feature buffer is read directly, and shallow convolution is performed only on the newly input current frame to extract edge and texture features.

5. The improved YOLO low-computing-power monitoring target recognition algorithm according to claim 4, characterized in that, Step S4 specifically includes the following steps: S41. Based on the feature reuse mechanism described in S3, an auxiliary detection mechanism based on short-term motion trends is introduced; for a continuous frame sequence reused in S3, a target tracking list is continuously maintained, specifically recording the position coordinates and displacement vectors of the target that has been successfully locked in the previous frame. S42. Assuming that the target in the monitored scene maintains uniform linear motion for a short period of time, based on the displacement vector recorded in S41, the expected position of the target in the current frame is calculated; the target center coordinates and displacement vector of the previous frame are added together to obtain the predicted center coordinates of the current frame, and a candidate region suggestion box is generated with the predicted center coordinates as the center. S43. When the target response weakens in the current frame due to the reduction of input resolution or extraction of only shallow features in S3, a feature enhancement strategy is initiated; the spatial position of the candidate region proposal box generated in S42 on the feature map is located, and the feature map amplitude within the candidate region is multiplied by an enhancement coefficient in the range of 1.5 to 3.

0. S44. The steps described above work in conjunction with the dynamic reasoning mechanism of S3. Feature enhancement is triggered when S3 determines that the calculation is in a static or shallow calculation mode, and automatically goes into hibernation during high-resolution full calculation.

6. The improved YOLO low-computing-power monitoring target recognition algorithm according to claim 5, characterized in that, Step S5 specifically includes the following steps: S51. Construct a resource monitoring thread for the edge device. The resource monitoring thread runs independently of the inference main process and periodically collects the real-time operating status of the chip. The collected indicators specifically include processor utilization, memory read / write bandwidth, and task queue length of the graphics computing unit. The three indicators are weighted and summed to generate a comprehensive computing power score. The weighting coefficients for processor utilization, memory read / write bandwidth, and task queue length of the graphics computing unit are set to 0.4, 0.3, and 0.3, respectively. S52. Based on the comprehensive computing power score generated in S51, dynamically adjust the execution strategy of the post-processing pipeline; when the comprehensive computing power score is lower than the preset threshold, enable the lightweight mode, simplify the candidate box screening process based on intersection-union ratio to directly extract the top K candidate boxes after sorting by confidence score, and temporarily lower the confidence threshold; when the comprehensive computing power score is higher than the preset threshold, resume frame-by-frame intersection-union comparison screening, and adopt a strict confidence threshold. S53. To address the candidate box scale differences caused by dynamic resolution switching in S3, a scale compensation coefficient is introduced in the post-processing stage. The corresponding scaling factor needs to be calculated based on the ratio between the current input resolution and the original input image resolution. Using the scaling factor, the coordinates of the candidate boxes generated at different resolutions are mapped back to the scale space of the original input image. S54. Refactor the post-processing code logic, merging bounding box decoding, scale compensation, and candidate box filtering into a kernel startup process for a single graphics calculation unit.

7. The improved YOLO low-computing-power monitoring target recognition algorithm according to claim 6, characterized in that, Step S6 specifically includes the following steps: S61. Before deploying the inference engine, perform a mixed precision conversion process. For the feature extraction structure after pruning in S1, simulate the low-bit integer operation characteristics of low-computing-power chips during the forward inference process. Before the weight parameters and activation values ​​enter the convolution kernel, artificially apply quantization noise. Then, retrain the full dataset. S62. Combine the static mask matrix generated in S2 and implement a non-uniform precision allocation strategy; count the importance of each position along the spatial dimension of the feature map, define the region marked with high weight by the static mask as the high-precision calculation region, and define the remaining region as the low-precision calculation region. S63. For the different precision regions defined above, implement differentiated memory management during the inference initialization phase; specify the input and output tensors of the computation layer corresponding to the high-precision computation region as higher-precision floating-point format; specify the tensors corresponding to the low-precision computation region as low-bit integer format; and allocate independent memory alignment blocks for these two types of tensors of different data types respectively. S64. Establish a closed-loop adjustment mechanism for quantization parameters and hardware load; based on the comprehensive computing power score generated in S5, dynamically fine-tune the quantization cutoff threshold and scaling factor of each layer; set constraints to limit the fine-tuning range to ±5% of the initial quantization parameters. S65, the steps mentioned above form a close upstream and downstream linkage with the structural compression of S1 and the hardware-aware scheduling of S5; S1 reduces the total amount of weight parameters that need to be quantized through pruning.

8. The improved YOLO low-computing-power monitoring target recognition algorithm according to claim 7, characterized in that, In step S14, the specific operations for assessing the importance of the channel include: Using the scaling factor output by the normalization layer as the channel importance weight coefficient, calculate the L1 norm or mean of the scaling factor; set the channel pruning rate, sort the scaling factor values ​​in ascending order, and select the channel with the smallest value as a redundant channel for removal; the removal operation includes setting the convolution kernel weight parameter of the redundant channel in the feature extraction structure to zero, and updating the channel index mapping table of subsequent calculation layers.

9. The improved YOLO low-computing-power monitoring target recognition algorithm according to claim 8, characterized in that, In step S34, the triggering logic for the cache validity rule specifically includes: A frame counter and a differential accumulator are set. The frame counter increments by 1 after each frame is processed. When the frame counter value reaches the upper limit of the range of 15 to 30 frames, a forced refresh signal is triggered. The differential accumulator accumulates the sum of the absolute values ​​of the pixel differences between adjacent frames. When the differential accumulator value exceeds the high resolution switching threshold, an instant refresh signal is triggered. After any refresh signal is triggered, the frame counter and the differential accumulator are reset, and the feature extraction structure described in S1 is called to re-execute the full feature extraction.

10. An improved YOLO low-computing-power monitoring target recognition algorithm according to claim 9, characterized in that, The specific implementation of the single graphics computing unit kernel startup process in step S54 includes: In the kernel startup function, a shared memory buffer is defined, and the original coordinate data generated by the bounding box decoding, the scaling parameters calculated by the scale compensation coefficient, and the confidence threshold for candidate box filtering are passed in as input parameters all at once. Inside the graphics computing unit, coordinate inverse normalization operation, scale space mapping operation, and confidence-based candidate box filtering operation are executed sequentially. The operation results are directly written to the output video memory, and intermediate results are not copied or synchronized between the processor and the graphics computing unit.

Images