Edge computing rail transit video AI intelligent analysis board card

Abstract

The application discloses an edge computing rail transit video AI intelligent analysis board card, and relates to the technical field of intelligent analysis board cards.The application solves the technical problems of disjunction between computing power scheduling and video content complexity and heterogeneous chip regulation homogenization.The application establishes a chip temperature resistance characteristic database containing junction temperature nominal upper limit, performance safety upper limit, service life constraint upper limit, early warning / critical temperature and maximum frequency reduction ratio through offline calibration.When the chip is in different temperature intervals, different frequency reduction ratio calculation methods are adopted, especially the use of thermal matching degree score to dynamically adjust the frequency reduction amplitude, which reduces the temperature range between chips, improves the overall energy efficiency ratio of the system, and proposes a thermal matching degree score formula.When the score is lower than the migration threshold, task migration is actively triggered instead of blind frequency reduction, which solves the technical problem of further performance deterioration caused by forced frequency reduction of low energy efficiency tasks, improves the GPU utilization rate of non-safety critical tasks, and ensures the real-time performance of safety critical tasks.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis board technology, specifically to an edge computing rail transit video AI intelligent analysis board. Background Technology

[0002] The edge computing rail transit video AI intelligent analysis system deploys intelligent analysis boards at the edge nodes of the trackside, platform and vehicle to process video streams in real time and realize key functions such as foreign object intrusion detection, passenger flow density statistics and passenger abnormal behavior recognition. In the existing technology, such boards usually integrate multiple heterogeneous computing units such as NPU, GPU, FPGA and CPU to balance energy efficiency and computing power.

[0003] However, existing technologies employ static or simple load-based scheduling strategies, which cannot dynamically adjust the working state of each computing unit according to the video content. This results in wasted power consumption in low-complexity scenarios and insufficient computing power in high-complexity scenarios. Furthermore, existing solutions use uniform temperature thresholds and frequency reduction for each chip, ignoring the fundamental differences in temperature resistance, thermal capacity, thermal time constant, and task thermal characteristics among NPUs, GPUs, CPUs, and FPGAs. This leads to premature frequency reduction for chips with large thermal capacity and frequent overheating of chips with small thermal capacity, resulting in low overall system energy efficiency. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an edge computing AI intelligent analysis board for rail transit videos, which solves the problems of disconnect between computing power scheduling and video content complexity, as well as the homogenization of heterogeneous chip control.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an edge computing rail transit video AI intelligent analysis board, comprising:

[0006] The video content evaluation module is used to extract keyframes from each video stream, calculate their motion vector entropy, target density index, and illumination change rate, and input them into a pre-trained random forest classifier, outputting extremely low, low, medium, and high complexity levels.

[0007] The resource allocation and processing module is used to perform differentiated computing power configuration according to the complexity level, and sequentially configure the NPU to be downclocked by 40% and frame-skipping to be overclocked by 110%, the GPU to be powered off and then full speed, and the FPGA to be turned off and then full speed to take over the convolutional layer in a stepped computing power allocation.

[0008] The heat dissipation adjustment processing module is used to collect sensor data and extract short-term, medium-term and long-term features through a deployed multi-point temperature sensor array. These features, along with the video complexity level, are input into a single-layer GRU network to predict the temperature in the next 10 seconds. The predicted temperature is combined with the actual temperature to identify the thermal load status, and a general control strategy is executed based on the thermal load status.

[0009] The differentiated collaborative control module is used to calculate the task-chip thermal matching score based on the offline calibrated chip temperature resistance database and task thermal characteristics, on the basis of a general strategy. Based on the chip temperature resistance database and the task thermal matching score, and combined with the collected real-time junction temperature comparison threshold, differentiated control is performed.

[0010] As a further aspect of the present invention, the method for calculating the motion vector entropy, target density index, and illumination change rate is as follows:

[0011] Running vector entropy: The number of feature points labeled in each frame of the image is N, and each feature point i corresponds to a velocity vector. =(u i v i ), the direction angle of the velocity vector =arctan(v i / u i The feature points are quantized into M intervals, and the percentage p of feature points falling in the k-th direction interval is calculated. k ,in c k This represents the number of samples falling into the k-th interval, according to the formula... The running vector entropy H is calculated. 熵 ;

[0012] Target density index: By running a lightweight target detection network, all targets N and their corresponding bounding boxes in the frame are detected, according to the formula... The target density index D is calculated, where W is the frame width, H is the frame height, and Aa is the area of ​​the bounding box of the a-th target.

[0013] Illumination change rate: The current frame I t The color space is converted from RGB to HSV for color statistics, and the luminance component V is extracted. The color histogram H of the luminance component of the frame is then calculated. t Simultaneously calculate the histogram difference between the current frame and the previous frame. Where B represents the number of intervals in the histogram, b is the interval label of the histogram, and according to the formula... The rate of change of illumination L was calculated. t ,in This is the normalization factor.

[0014] As a further aspect of the present invention, the resource allocation processing module performs differentiated computing power configuration based on complexity level in the following manner:

[0015] Very low level: NPU frequency reduced to 40% and frame rate cut by 1 / 2, some cores shut down; GPU powered off; FPGA shut down or only format conversion; only critical event detection;

[0016] Low level: NPU running at 70%, full-frame processing; GPU powered off; FPGA with motion deblurring enabled; standard object detection and tracking.

[0017] Mid-range: NPU at full frequency; GPU wake-up and half frequency; FPGA-assisted feature extraction; target detection + classification + initial behavior screening;

[0018] High-level: NPU overclocked to 110%; GPU at full speed; FPGA at full speed and taking over some convolutional layers; NPU initial inspection + GPU fine classification + FPGA key layer acceleration.

[0019] As a further aspect of the present invention, the short-term characteristics include the instantaneous temperature change rate within the most recent second and the temperature difference between each chip.

[0020] Medium-term characteristics include the temperature moving average and standard deviation over the most recent 30 seconds; long-term characteristics include the temperature change pattern over the most recent 5 minutes.

[0021] As a further aspect of the present invention, the specific method for identifying the thermal load state is as follows:

[0022] If the predicted temperature is less than the first preset temperature and the actual temperature is less than the first preset temperature, then it is determined to be a normal state.

[0023] If the predicted temperature is between the first preset temperature and the second preset temperature, it is determined to be a warning state;

[0024] If the predicted temperature is greater than the second preset temperature or the actual temperature is greater than the second preset temperature, then it is determined to be a critical state.

[0025] The first preset temperature is 72°C to 83°C, depending on the chip type, and the second preset temperature is 80°C to 95°C, depending on the chip type.

[0026] As a further aspect of the present invention, the method of executing the general control strategy according to the thermal load state is as follows:

[0027] Under normal conditions, the fan or solid-state heatsink runs at its lowest speed and does not actively reduce the frequency.

[0028] In alarm mode, the fan duty cycle increases linearly to 60% to 80%, and the NPU frequency is reduced to 80% of the target value for the current complexity level. If the GPU is awakened, its frequency is limited to below 50%.

[0029] In critical conditions, the fan or solid-state heatsink runs at 100% duty cycle, the NPU frequency drops to 30% to 50% of the nominal frequency, the GPU is forced to hibernate, and all non-safety-critical tasks are suspended.

[0030] As a further aspect of the present invention, the method for scoring the computational task-chip thermal compatibility is as follows:

[0031] According to the formula Calculate the thermal fit score S for each task and chip combination. 热 , where P s P represents the inference speed of the task on this chip. 功耗 W represents the average power consumption generated when the task runs on this chip. sens This indicates the chip's temperature sensitivity coefficient, T. re This indicates the temperature difference between the chip and its current performance limit, in T. 上限 This indicates the upper limit of the chip's performance safety temperature. This is the scaling factor.

[0032] As a further aspect of the present invention, the method of differential regulation by combining the collected real-time junction temperature comparison threshold is as follows:

[0033] If the current junction temperature is below the lifetime constraint limit, maintain the chip's current operating frequency;

[0034] If the current junction temperature reaches the lifetime constraint limit but does not exceed the recommended warning temperature, the frequency reduction ratio will be adjusted according to the thermal matching score.

[0035] If the junction temperature exceeds the recommended warning temperature but does not reach the recommended critical temperature, the frequency will be reduced by half of the highest frequency reduction ratio for the corresponding chip.

[0036] If the junction temperature reaches or exceeds the recommended critical temperature, the highest frequency reduction ratio of the corresponding chip will be directly activated, and the GPU will directly enter a sleep state. The frequency will be gradually restored after the junction temperature drops back to a safe range.

[0037] As a further aspect of the present invention, the method of slightly adjusting the frequency reduction ratio based on the thermal matching score is as follows:

[0038] Within the target temperature range, the formula for calculating the frequency reduction ratio F is: , where F base k is the upper limit of the baseline down-frequency ratio. cap This is the normalized coefficient for the chip's thermal capacity. The coefficient of temperature change. Given the current temperature change rate, frequency reduction is performed based on the calculated frequency reduction ratio F.

[0039] As a further aspect of the present invention, the conditions for gradually restoring the frequency of the GPU after it has gone into sleep mode are as follows:

[0040] When the junction temperature is 5°C lower than the recommended warning temperature for 5 seconds and the temperature change rate for 5 consecutive samples is less than 0.5°C / s, the GPU is woken up at 30% of the nominal frequency. After running for 5 seconds, if the temperature does not exceed the warning value, the frequency is increased by 10% every 200ms until the target frequency is reached.

[0041] This invention provides an edge computing AI intelligent analysis board for rail transit video. Compared with existing technologies, it has the following advantages:

[0042] This invention establishes a chip temperature resistance database through offline calibration, which includes the nominal upper limit of junction temperature, the upper limit of performance safety, the upper limit of lifespan constraint, the warning / critical temperature, and the maximum frequency reduction ratio. It also introduces a chip thermal capacity normalization coefficient. When the chip is in different temperature ranges, different frequency reduction ratio calculation methods are used. In particular, the frequency reduction amplitude is dynamically adjusted by using thermal matching score to reduce the temperature difference between chips and improve the overall energy efficiency ratio of the system.

[0043] This invention proposes a thermal matching score formula that integrates task execution speed, average power consumption, chip temperature sensitivity coefficient, remaining heat capacity, and task priority. When the score is lower than the migration threshold, task migration is actively triggered instead of blindly reducing the frequency. This solves the technical problem of forced frequency reduction of low-energy-efficiency tasks leading to further performance degradation, improves the GPU utilization of non-safety-critical tasks, and ensures the real-time performance of safety-critical tasks. Attached Figure Description

[0044] Figure 1 This is a functional diagram of the edge computing rail transit video AI intelligent analysis board of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] First Embodiment

[0047] Please see Figure 1 This application provides an edge computing rail transit video AI intelligent analysis board, including:

[0048] The video content evaluation module is used to extract keyframes from each video stream, calculate their motion vector entropy, target density index, and illumination change rate, and input them into a pre-trained random forest classifier, outputting extremely low, low, medium, and high complexity levels.

[0049] The resource allocation and processing module is used to perform differentiated computing power configuration according to the complexity level, and sequentially configure the NPU to be downclocked by 40% and frame-skipping to be overclocked by 110%, the GPU to be powered off and then full speed, and the FPGA to be turned off and then full speed to take over the convolutional layer in a stepped computing power allocation.

[0050] The heat dissipation adjustment processing module is used to collect sensor data and extract short-term, medium-term and long-term features through a deployed multi-point temperature sensor array. These features, along with the video complexity level, are input into a single-layer GRU network to predict the temperature in the next 10 seconds. The predicted temperature is combined with the actual temperature to identify the thermal load status, and a general control strategy is executed based on the thermal load status.

[0051] The differentiated collaborative control module is used to calculate the task-chip thermal matching score based on the offline calibrated chip temperature resistance database and task thermal characteristics, on the basis of a general strategy. Based on the chip temperature resistance database and the task thermal matching score, and combined with the collected real-time junction temperature comparison threshold, differentiated control is performed.

[0052] Second Embodiment

[0053] As a second embodiment of the present invention, the difference from the first embodiment is as follows:

[0054] The video content evaluation module continuously extracts keyframes from each video stream through the main control CPU and calculates the running vector entropy, target density index, and illumination change rate of the keyframes.

[0055] Running vector entropy: The number of feature points in each frame of the image is denoted as N, and each feature point i corresponds to a velocity vector. =(u i v i ), the direction angle of the velocity vector =arctan(v i / u i The feature points are quantized into M intervals, and the percentage p of feature points falling in the k-th direction interval is calculated. k ,in c k This represents the number of samples falling into the k-th interval, and Further according to the formula The running vector entropy H is calculated. 熵 ;

[0056] Target density index: Let the total number of pixels in a video frame be S = W × H, where W is the frame width and H is the frame height. By running a lightweight target detection network, all targets N and their corresponding bounding boxes in the frame are detected. The area of ​​the bounding box of the a-th target is obtained as Aa. According to the formula... The target density index D is calculated;

[0057] Illumination change rate: The current frame I t The color space is converted from RGB to HSV for color statistics, and the luminance component V is extracted. The color histogram H of the luminance component of the frame is then calculated. t Simultaneously calculate the histogram difference between the current frame and the previous frame. Where B represents the number of intervals in the histogram, b is the interval label of the histogram, and according to the formula... The rate of change of illumination L was calculated. t ,in Normalization factor;

[0058] The obtained running vector entropy, target density index, and illumination change rate are input into a pre-trained random forest classifier, which contains 20 decision trees, each with a maximum depth of 3, and outputs a complexity level C, where C∈[L1, L2, L3, L4], where L1 represents very low, L2 represents low, L3 represents medium, and L4 represents high.

[0059] At the same time, the obtained video complexity level is transmitted to the resource allocation and processing module.

[0060] The resource allocation and processing module allocates computing resources according to the complexity level of the video. For extremely low complexity levels, the NPU is downclocked to 40% of its nominal frequency and only processes 1 / 2 of the frame extraction. Some computing cores are shut down, the GPU is powered off through a load switch, the FPGA is shut down or only performs simple format conversion, and the inference mode is only used for critical event detection.

[0061] For low-level complexity, the NPU runs at 70% of its nominal frequency for full-frame processing, the GPU is powered off, the FPGA is powered on for motion deblurring preprocessing, and the inference mode is used for regular target detection and tracking.

[0062] For medium-level complexity, the NPU runs at 100% of its nominal frequency, the GPU wakes up from deep sleep and runs at 50% of its frequency, the FPGA is turned on and performs auxiliary feature extraction, and the inference mode is object detection + classification + behavior screening.

[0063] For high-level complexity, the NPU is overclocked to 110%, the GPU runs at full speed, the FPGA runs at full speed, and takes over some convolutional layers. The inference mode is NPU initial detection + GPU fine classification + FPGA acceleration of key layers.

[0064] Simultaneously read temperature sensor data in real time. If the junction temperature of the computing unit exceeds the preset temperature, a migration strategy is triggered. If the complexity level is L2 or L3, some computing tasks are migrated to the NPU or FPGA, and the GPU frequency is reduced simultaneously. If the level is L4 and the GPU is already running at full speed, some non-real-time tasks are actively transferred to the cloud or adjacent edge nodes, and the heat dissipation strategy is adjusted.

[0065] The heat dissipation adjustment module is used to adjust the heat dissipation according to the real-time temperature of the board. Specifically, a multi-point temperature sensor array is deployed at the key heat source locations of the edge computing board to collect sensor data. At the same time, short-term, medium-term, and long-term features are extracted from the sensor data. The short-term features represent the instantaneous temperature change rate and the temperature difference between each chip in the last second. The medium-term features represent the temperature sliding average and standard deviation in the last 30 seconds. The long-term features represent the temperature change pattern in the last 5 minutes.

[0066] Based on the extracted three types of temperature features, they are input together with the video content complexity level into a lightweight gated recurrent unit (GRU) network to predict the temperature for the next 10 seconds. The lightweight GRU network has a single-layer GRU structure with 32 hidden layers, and the output dimension corresponds to the predicted future temperature value for each chip. It is pre-trained in the cloud using data from various operating scenarios, and the network parameters are fine-tuned online every 24 hours using actual operating data. At the same time, the actual temperature is acquired, and the two are combined to identify the thermal load status.

[0067] If the predicted temperature is less than the first preset temperature and the actual temperature is less than the first preset temperature, it is determined to be in a normal state. If the predicted temperature is between the first preset temperature and the second preset temperature, it is determined to be in a warning state. If the predicted temperature is greater than the second preset temperature or the actual temperature is greater than the second preset temperature, it is determined to be in a critical state. The first preset temperature is 72℃~83℃ depending on the chip type, and the second preset temperature is 80℃~95℃ depending on the chip type.

[0068] At the same time, different control strategies are implemented according to different thermal load states to generate a general control strategy. Specifically, for normal conditions, the fan or solid-state heat dissipation module runs at the lowest speed and relies on the heat dissipation fins and phase change materials to maintain thermal balance. The phase change materials remain solid and do not actively reduce the frequency. The frequency is completely scheduled according to the video complexity level.

[0069] In the event of an alert, the fan duty cycle is linearly increased to 60% to 80%, and the fan duty cycle is calculated as follows: T 实际The actual temperature is indicated. The NPU frequency is reduced to 80% of the target value for the current complexity level. If the GPU is awake, its frequency is limited to below 50%. If the GPU temperature exceeds the second preset temperature, it is forced into deep sleep. At the same time, non-urgent inference tasks are migrated from the GPU to the NPU. If the NPU load is too high, it is migrated to the cloud to assist inference.

[0070] In critical situations, the fan or solid-state cooling module operates at 100% duty cycle to achieve maximum heat dissipation capacity, the NPU frequency drops to 30% to 50% of the nominal frequency, the GPU is forced to hibernate, and all non-safety-critical tasks are suspended.

[0071] For example, in a scenario where a sudden surge in passenger flow occurs on a platform, causing the video complexity level to jump from L2 to L4, at t=0s, the complexity assessment module detects L4, the scheduler wakes up the GPU, and the NPU is overclocked to 110%. At this point, the hot load status changes from normal to warning.

[0072] At t+10ms, the RNN prediction model outputs that the temperature will rise from 72℃ to 83℃ in the next 5 seconds, and the thermal controller performs pre-cooling: the fan duty cycle jumps from 30% to 90%.

[0073] At t+200ms, the actual NPU temperature reached 78℃ and the GPU temperature reached 70℃. The system remained in a warning state and limited the GPU frequency to 50% to avoid overheating.

[0074] At t+2s, the large passenger flow continued, and the NPU temperature rose to 82℃, but did not reach the critical point. The fan maintained 90% speed, the phase change material began to melt and absorb heat, and the rate of temperature rise slowed down.

[0075] At t+10, passenger flow decreased, the complexity level dropped to L2, the GPU went into sleep mode, the NPU frequency dropped to 70%, and the temperature began to decrease.

[0076] At t+15s, the temperature dropped to 72℃, and the predicted temperature was below 75℃. The status was switched back to normal, and the fan duty cycle gradually decreased to 40%.

[0077] Third Embodiment

[0078] As a third embodiment of the present invention, it is implemented based on the second embodiment, and the difference from the second embodiment is as follows:

[0079] The differentiated collaborative control module, based on the general control strategy generated in the second embodiment, performs differentiated adjustments to the temperature resistance and heat dissipation capabilities of different chips. The specific adjustment method is as follows:

[0080] System calibration was performed on each type of chip through offline testing. The chips included NPU, GPU, CPU, and FPGA, and a temperature resistance database was established, as shown in the table below:

[0081]

[0082] In the table, the nominal junction temperature upper limit indicates the maximum allowable junction temperature specified in the chip datasheet; the performance safety upper limit indicates that performance will significantly decrease after exceeding this temperature, but the chip can still operate safely; and the lifetime constraint upper limit indicates that, based on the Arrhenius lifetime model, long-term operation above this temperature will significantly shorten the expected lifespan of the chip, and runtime constraints are required.

[0083] Simultaneously, offline testing was used to calibrate the thermal characteristics of each type of chip during operation. These thermal characteristics included peak power consumption, average power consumption, power consumption curve time pattern, and temperature sensitivity coefficient. Peak power consumption represents the maximum instantaneous power consumption during task execution, average power consumption represents the average power consumption within the task execution cycle, and the temperature sensitivity coefficient represents the increase in inference latency caused by a 1°C increase in temperature. The results were then calculated using the formula... Calculate the thermal fit score S for each task and chip combination. 热 Furthermore, the parameters in the above formula are normalized values, where P s P represents the inference speed of the task on this chip. 功耗 W represents the average power consumption generated when the task runs on this chip. sens This indicates the chip's temperature sensitivity coefficient, T. re This indicates the temperature difference between the chip and its current performance limit, in T. 上限 This indicates the upper limit of the chip's performance safety temperature.

[0084] The current junction temperature of each chip is collected synchronously in real time. First, it is compared with the temperature thresholds of each level in the temperature resistance characteristic data table. Then, the control range under the current operating condition is calculated by combining the calibrated thermal characteristics of each type of chip with the matching situation of the current task.

[0085] If the current junction temperature is below the lifetime constraint limit, the chip's current operating frequency will be maintained without frequency reduction. If the current junction temperature reaches the lifetime constraint limit but does not exceed the recommended warning temperature, the frequency reduction ratio will be slightly adjusted based on the thermal matching score. The specific adjustment method is as follows:

[0086] Within the target temperature range, the formula for calculating the frequency reduction ratio F is: , where F base k is the upper limit of the baseline down-frequency ratio. cap This is the normalized coefficient for the chip's thermal capacity. The coefficient of temperature change. Given the current rate of temperature change, frequency reduction is performed based on the calculated frequency reduction ratio F.

[0087] If the junction temperature exceeds the recommended warning temperature but does not reach the recommended critical temperature, the frequency will be reduced by half of the highest frequency reduction ratio for the corresponding chip.

[0088] If the junction temperature reaches or exceeds the recommended critical temperature, the highest frequency reduction ratio of the corresponding chip will be directly activated, and the GPU will directly enter a sleep state. The frequency will be gradually restored after the junction temperature drops back to a safe range. The specific conditions are as follows: when the junction temperature is 5°C lower than the recommended warning temperature for 5 seconds and the temperature change rate of 5 consecutive samples is less than 0.5°C / s, the GPU will be woken up at 30% of the nominal frequency. After running for 5 seconds, if the temperature does not exceed the warning value, the frequency will be increased by 10% every 200ms until the target frequency is reached.

[0089] Fourth embodiment

[0090] As a fourth embodiment of the present invention, the focus is on combining the implementation processes of all the above embodiments.

[0091] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0092] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An edge computing rail transit video AI intelligent analysis board, characterized in that, include: The video content evaluation module is used to extract keyframes from each video stream, calculate their motion vector entropy, target density index, and illumination change rate, and input them into a pre-trained random forest classifier, outputting extremely low, low, medium, and high complexity levels. The resource allocation and processing module is used to perform differentiated computing power configuration according to the complexity level, and sequentially configure the NPU to be downclocked by 40% and frame-skipping to be overclocked by 110%, the GPU to be powered off and then full speed, and the FPGA to be turned off and then full speed to take over the convolutional layer in a stepped computing power allocation. The heat dissipation adjustment processing module is used to collect sensor data and extract short-term, medium-term and long-term features through a deployed multi-point temperature sensor array. These features, along with the video complexity level, are input into a single-layer GRU network to predict the temperature in the next 10 seconds. The predicted temperature is combined with the actual temperature to identify the thermal load status, and a general control strategy is executed based on the thermal load status. The differentiated collaborative control module is used to calculate the task-chip thermal matching score based on the offline calibrated chip temperature resistance database and task thermal characteristics, on the basis of a general strategy. Based on the chip temperature resistance database and the task thermal matching score, and combined with the collected real-time junction temperature comparison threshold, differentiated control is performed.

2. The edge computing rail transit video AI intelligent analysis board according to claim 1, characterized in that, The method for calculating the motion vector entropy, target density index, and rate of change of illumination is as follows: Running vector entropy: The number of feature points labeled in each frame of the image is N, and each feature point i corresponds to a velocity vector. =(u i v i ), the direction angle of the velocity vector =arctan(v i / u i The feature points are quantized into M intervals, and the percentage p of feature points falling in the k-th direction interval is calculated. k ,in c k This represents the number of samples falling into the k-th interval, according to the formula... The running vector entropy H is calculated. 熵 ; Target density index: By running a lightweight target detection network, all targets N and their corresponding bounding boxes in the frame are detected, according to the formula... The target density index D is calculated, where W is the frame width, H is the frame height, and Aa is the area of ​​the bounding box of the a-th target. Illumination change rate: The current frame I t The color space is converted from RGB to HSV for color statistics, and the luminance component V is extracted. The color histogram H of the luminance component of the frame is then calculated. t Simultaneously calculate the histogram difference between the current frame and the previous frame. Where B represents the number of intervals in the histogram, b is the interval label of the histogram, and according to the formula... The rate of change of illumination L was calculated. t ,in This is the normalization factor.

3. The edge computing rail transit video AI intelligent analysis board according to claim 1, characterized in that, The resource allocation processing module performs differentiated computing power configuration based on complexity level as follows: Very low level: NPU frequency reduced to 40% and frame rate cut by 1 / 2, some cores shut down; GPU powered off; FPGA shut down or only format conversion; only critical event detection; Low level: NPU running at 70%, full-frame processing; GPU powered off; FPGA with motion deblurring enabled; standard object detection and tracking. Mid-range: NPU at full frequency; GPU wake-up and half frequency; FPGA-assisted feature extraction; target detection + classification + initial behavior screening; High-level: NPU overclocked to 110%; GPU at full speed; FPGA at full speed and taking over some convolutional layers; NPU initial inspection + GPU fine classification + FPGA key layer acceleration.

4. The edge computing rail transit video AI intelligent analysis board according to claim 1, characterized in that, The short-term characteristics include the instantaneous temperature change rate within the last second and the temperature difference between each chip. Medium-term characteristics include the temperature moving average and standard deviation over the most recent 30 seconds; long-term characteristics include the temperature change pattern over the most recent 5 minutes.

5. The edge computing rail transit video AI intelligent analysis board according to claim 1, characterized in that, The specific method for identifying the thermal load status is as follows: If the predicted temperature is less than the first preset temperature and the actual temperature is less than the first preset temperature, then it is determined to be a normal state. If the predicted temperature is between the first preset temperature and the second preset temperature, it is determined to be a warning state; If the predicted temperature is greater than the second preset temperature or the actual temperature is greater than the second preset temperature, then it is determined to be a critical state. The first preset temperature is 72°C to 83°C, depending on the chip type, and the second preset temperature is 80°C to 95°C, depending on the chip type.

6. The edge computing rail transit video AI intelligent analysis board according to claim 1, characterized in that, The method for implementing the general control strategy based on the thermal load state is as follows: Under normal conditions, the fan or solid-state heatsink runs at its lowest speed and does not actively reduce the frequency. In alarm mode, the fan duty cycle increases linearly to 60% to 80%, and the NPU frequency is reduced to 80% of the target value for the current complexity level. If the GPU is awakened, its frequency is limited to below 50%. In critical conditions, the fan or solid-state heatsink runs at 100% duty cycle, the NPU frequency drops to 30% to 50% of the nominal frequency, the GPU is forced to hibernate, and all non-safety-critical tasks are suspended.

7. The edge computing rail transit video AI intelligent analysis board according to claim 1, characterized in that, The method for scoring the computation task-chip thermal compatibility is as follows: According to the formula Calculate the thermal fit score S for each task and chip combination. 热 , where P s P represents the inference speed of the task on this chip. 功耗 W represents the average power consumption generated when the task runs on this chip. sens This indicates the chip's temperature sensitivity coefficient, T. re This indicates the temperature difference between the chip and its current performance limit, in T. 上限 This indicates the upper limit of the chip's performance safety temperature. This is the scaling factor.

8. The edge computing rail transit video AI intelligent analysis board according to claim 1, characterized in that, The method of differential regulation by combining the collected real-time junction temperature comparison threshold is as follows: If the current junction temperature is below the lifetime constraint limit, maintain the chip's current operating frequency; If the current junction temperature reaches the lifetime constraint limit but does not exceed the recommended warning temperature, the frequency reduction ratio will be adjusted according to the thermal matching score. If the junction temperature exceeds the recommended warning temperature but does not reach the recommended critical temperature, the frequency will be reduced by half of the highest frequency reduction ratio for the corresponding chip. If the junction temperature reaches or exceeds the recommended critical temperature, the highest frequency reduction ratio of the corresponding chip will be directly activated, and the GPU will directly enter a sleep state. The frequency will be gradually restored after the junction temperature drops back to a safe range.

9. The edge computing rail transit video AI intelligent analysis board according to claim 8, characterized in that, The method of slightly adjusting the frequency reduction ratio based on the thermal matching score is as follows: Within the target temperature range, the formula for calculating the frequency reduction ratio F is: , where F base k is the upper limit of the baseline down-frequency ratio. cap This is the normalized coefficient for the chip's thermal capacity. The coefficient of temperature change. Given the current temperature change rate, frequency reduction is performed based on the calculated frequency reduction ratio F.

10. The edge computing rail transit video AI intelligent analysis board according to claim 8, characterized in that, The conditions for the GPU to gradually restore its frequency after hibernation are: When the junction temperature is 5°C lower than the recommended warning temperature for 5 seconds and the temperature change rate for 5 consecutive samples is less than 0.5°C / s, the GPU is woken up at 30% of the nominal frequency. After running for 5 seconds, if the temperature does not exceed the warning value, the frequency is increased by 10% every 200ms until the target frequency is reached.

Images