A spatio-temporal feature fusion method for platform environment analysis

Abstract

This invention relates to the field of neural network computation graph adaptive configuration and artificial intelligence optimized operating system technology, and discloses a spatiotemporal feature fusion method for station environment analysis, including: extracting the rate of change of spatiotemporal tensor residual scalar generated by the accumulation of spatiotemporal feature tensors with consecutive frames to adjust the weight update of the kernel scheduling tree of the real-time microkernel operating system, and calculating and generating a core ready queue impedance index based on the utilization rate of the multi-core computing power cores and the bus bandwidth occupancy rate of the operating system; when the index is greater than the safe impedance threshold, triggering the cache fault tolerance control flow, reducing the scheduling priority of the kernel scheduling thread and injecting a step-by-step penalty operator to apply dynamic pruning of the spatial dimension. This invention breaks the decoupling limitation between the computation model and resource scheduling, enabling the operating system to perceive the causal characteristics of the algorithm topology, reconstructing the task pipeline under high load to offset latency jitter, ensuring deterministic computation latency, and avoiding system crashes induced by resource inversion.

Description

Technical Field

[0001] This invention belongs to the field of neural network computation graph adaptive configuration and artificial intelligence optimized operating system technology, and particularly relates to a spatiotemporal feature fusion method for station environment analysis. Background Technology

[0002] Currently, in the design of computer systems based on specific computational models, the heterogeneous resource management of advanced perceptual computing pipelines relies heavily on the real-time operating system scheduling mechanism. By rationally arranging task thread priorities at the kernel layer, the periodic and stable progress of multi-path high-dimensional spatiotemporal feature fusion computation is maintained. This establishes a dynamic resource distribution channel required for algorithm model graph structure computation within the entire computing platform firmware architecture. When the computer system faces the impact of high-density burst video stream perception, the convergence of multiple high-dimensional tensors in the spatial dimension and the accumulation of continuous frame features in the temporal dimension cause the hardware bus to generate a huge data throughput load instantaneously. The time-sharing scheduling mechanism of the operating system is structurally decoupled from the causal topology of the computational model and lacks the control capability to identify the spatiotemporal coupling characteristics between operators within the computation graph. This leads to frequent task inversion and a large amount of context switching dissipation when computing resources face peak burst non-steady loads, causing a physical conflict between the precision of perceptual computing and the determinism of system latency control, resulting in data backlog and system response lag.

[0003] To alleviate performance degradation caused by excessive consumption of computing resources, simply relying on the conventional path of expanding general-purpose cache capacity or increasing computing cores not only leads to increased data bus synchronization barriers and interlocking overhead between computing units, but also fails to eliminate control blind spots caused by the lack of awareness of the internal state of the model by the resource management unit. Hardware architecture improvements have limitations, and specific resource scheduling and control methods at the software level also have shortcomings. For example, Chinese invention patent application CN108694689A discloses a neural network scheduling mechanism that uses a kernel-mode driver and is based on preset priorities and... Submitting client type scheduling for hardware resource access is effective for steady-state general computing, but its technical approach relies on the underlying premise that the workload characteristics distribution is known and relatively constant. When applied to high-density, non-steady-state, burst video stream sensing conditions such as urban rail transit platforms, the scheduling logic based on static preset attributes is detached from the real-time dynamic operation state of the internal operators of the neural network. It cannot perceive the transient distortion and causal topology generated by the intersection of high-dimensional spatiotemporal feature streams in the spatiotemporal dimensions. It is prone to hardware bus computing power channel blockage due to the mismatch between static priority configuration and actual dynamic burst load, resulting in task priority inversion and system latency jitter.

[0004] Therefore, the technical problem to be solved by this invention is how to enable the real-time operating system kernel to directly obtain the spatiotemporal multidimensional data features of the current neural network computing model when using a spatiotemporal feature fusion method for station environment analysis, and to achieve adaptive architecture pruning and adaptive deterministic delay control of the neural network computing graph through the coordinated linkage of the kernel scheduling tree and heterogeneous task queues when the core hardware bus is overloaded. Summary of the Invention

[0005] This invention aims to solve the problem that real-time operating system kernels, due to their independent structure, cannot directly obtain the spatiotemporal multidimensional data characteristics of the computational model, and cause task inversion and latency jitter when the bus is overloaded.

[0006] In this technical solution, a spatiotemporal feature fusion method for platform environment analysis includes the following steps:

[0007] Step S1: Obtain the spatiotemporal feature tensor generated by the intersection of the spatial and temporal dimensions of the multidimensional sensing feature data stream corresponding to the platform environment input channel, and extract the rate of change of the spatiotemporal tensor residual scalar generated by the accumulation of the spatiotemporal feature tensor with consecutive frames during runtime.

[0008] Step S2: Adjust the weight update of the kernel scheduling tree of the real-time microkernel operating system according to the rate of change of the spatiotemporal tensor residual scalar to reconstruct the instruction flow orchestration, and use the utilization rate of the multi-core computing power core of the operating system and the bus bandwidth occupancy rate as weighting factors to aggregate calculations and generate the core ready queue impedance index that characterizes the degree of blockage of the hardware bus computing power channel.

[0009] Step S3: When the impedance index of the core ready queue is greater than the safety impedance threshold, the kernel scheduling unit triggers the time window decoupled cache fault-tolerant control flow, issues a kernel overwrite instruction to lower the scheduling priority of the kernel scheduling thread corresponding to a spatiotemporal feature fusion method for station environment analysis in the physical bus bandwidth by 3 scheduling levels, and simultaneously injects a step-by-step penalty operator into the computation graph corresponding to the spatiotemporal feature tensor in situ, closes the spatial convolution calculation branch of the secondary sensing feature input graph channel branch, and applies spatial dimension dynamic pruning to the computation graph in the spatial feature extraction step to stabilize the input tensor size within the channel range corresponding to the predetermined basic safety feature subset.

[0010] Preferably, the extraction of the scalar change rate of the spatiotemporal tensor residual in step S1 includes the following sub-steps: Step S11, calculate the tensor difference between the current frame spatiotemporal feature tensor and the feature vector copy of the previous cycle in the feature buffer to obtain the spatiotemporal tensor residual; Step S12, calculate the scalar magnitude of the spatiotemporal tensor residual and divide it by the frame interval time to generate the scalar change rate of the spatiotemporal tensor residual; Step S2, adjusting the weight update of the kernel scheduling tree of the real-time microkernel operating system includes: when the scalar change rate of the spatiotemporal tensor residual increases, proportionally increase the scheduling weight of the kernel scheduling thread in the kernel scheduling tree of the real-time microkernel operating system.

[0011] Preferably, step S3, which involves injecting a stepped penalty operator into the computation graph corresponding to the spatiotemporal feature tensor in situ, includes the following sub-steps: Step S31, obtaining the core task correlation degree corresponding to each secondary sensing feature input graph channel branch; Step S32, classifying each secondary sensing feature input graph channel branch into levels based on the core task correlation degree, and separating the key sensing input graph channel branches and the secondary sensing feature input graph channel branches; Step S33, when the impedance index of the core ready queue is in different overload level ranges, matching and closing the spatial feature extraction of the corresponding number of secondary sensing feature input graph channel branches through the stepped penalty operator.

[0012] Preferably, the cache fault-tolerant control flow that triggers time window decoupling in step S3 includes the following sub-steps: Step S34, when the multi-path spatial feature stream generates non-uniform transmission delay in the physical bus bandwidth transmission, interrupt the time accumulation calculation of the current time window; Step S35, forcibly call the cached feature vector copy of the previous cycle in the feature cache area through a zero-order hold action; Step S36, use the feature vector copy of the previous cycle to replace the multi-dimensional sensing feature data stream that has generated non-uniform transmission delay, and participate in the time accumulation calculation of the current cycle.

[0013] Preferably, after step S3, there is a step to reduce core sensing latency by releasing physical bus bandwidth due to closing the secondary sensing feature input graph channel branch: step S37, centrally schedule the multi-core computing power core of the operating system on the converged basic security feature subset to specifically ensure the deterministic computation latency of the core analysis task.

[0014] Preferably, the kernel scheduling tree of the real-time microkernel operating system adopts a red-black tree structure to organize kernel-mode threads. The weight update of adjusting the kernel scheduling tree of the real-time microkernel operating system in step S2 includes: recalculating the virtual running time of the kernel scheduling thread in the red-black tree structure according to the updated weight, so as to adjust its position in the red-black tree structure.

[0015] Preferably, step S3, which lowers the scheduling priority of the kernel scheduling thread corresponding to a spatiotemporal feature fusion method for station environment analysis in the physical bus bandwidth by 3 scheduling levels, includes: removing the kernel scheduling thread from the real-time scheduling queue of the real-time microkernel operating system and attaching it to the tail of a predetermined low-priority linked list of the time-sharing cyclic scheduling queue.

[0016] Preferably, the safety impedance threshold is a bus crash critical impedance scale determined based on hardware bus dissipation load data under sudden high throughput characteristic flow load conditions in historical periods.

[0017] Preferably, in step S3, converging the input tensor size to a predetermined subset of basic security features includes: after triggering the buffer fault-tolerant control flow, stabilizing the input tensor size within the channel range corresponding to the predetermined subset of basic security features, discarding the spatial convolutional feature tensor channels of the secondary sensing feature input map channel branches, and retaining only the spatial convolutional feature tensor channels containing the core analysis protection region.

[0018] Compared with existing technologies, the spatiotemporal feature fusion method for platform environment analysis proposed in this invention has the following advantages:

[0019] 1. In the spatiotemporal feature fusion of platform environment analysis, the spatiotemporal tensor residual scalar change rate is extracted by the state acquisition unit and the adaptive activation threshold is determined. This enables the system kernel to have a mechanism to dynamically adjust the scheduling tree weights based on the computational model operation state. When the system determines that the current working condition is in a steady state, the kernel scheduling unit directly modifies the scheduling tree logic of the current cycle, stops the spatial convolution operation of multi-path high-dimensional tensors, and uses the direct memory access mechanism to input the copy of the historical feature tensor of the previous cycle stored in the dedicated continuous physical memory page to the time feature accumulation operator. This completes the zero-order hold and logical yield of the computation pipeline, eliminating the idle thread scheduling loss caused by high-frequency polling under general time-sharing scheduling.

[0020] 2. By linking the task backlog length of the current heterogeneous computing unit core ready queue in the kernel register with the system data bus bandwidth utilization, a weighted calculation is performed to generate a core ready queue impedance index that reflects the degree of computing channel blockage. Based on this impedance index, when hardware resources face sudden overload, the kernel scheduling unit issues a kernel overriding instruction to lower the scheduling priority of the thread corresponding to the method by 3 levels. This not only cuts off the transient exclusive use of the bus by high-traffic spatial characteristic computing tasks, but also triggers the adaptive instruction flow reconstruction of the computing timing pipeline, realizing flexible allocation of multi-core computing power cores and bus bandwidth, and avoiding system-level avalanche caused by resource inversion of computing nodes.

[0021] 3. In response to high load overload conditions, the kernel scheduling unit uses the impedance index of the core ready queue as a hierarchical control criterion and injects a stepped penalty operator into the computation graph in situ. This operator forcibly shuts down the spatial feature extraction of some non-critical camera branches, and converges the processing tensor dimension required by the computation model to a subset of basic security features. In conjunction with the feature buffer separated by the time window, when non-uniform transmission delays occur in multiple spatial feature streams, the zero-order hold action forcibly calls the previous cycle feature vector copy in the buffer to participate in the time accumulation calculation. This enables the perception system to obtain adaptive adjustment capability to external sudden large traffic impacts with low resource cost, and ensures the deterministic latency of core security tasks. Attached Figure Description

[0022] Figure 1 This is a flowchart of the spatiotemporal feature fusion method for impedance discrimination and fault-tolerant control of the present invention;

[0023] Figure 2 This is a state diagram of the underlying scheduling mechanism of the spatiotemporal feature fusion method of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0025] A spatiotemporal feature fusion method for platform environment analysis includes the following steps:

[0026] Step S1: Obtain the spatiotemporal feature tensor generated by the intersection of the spatial and temporal dimensions of the multidimensional sensing feature data stream corresponding to the platform environment input channel, and extract the rate of change of the spatiotemporal tensor residual scalar generated by the accumulation of the spatiotemporal feature tensor with consecutive frames during runtime.

[0027] Step S2: Adjust the weight update of the kernel scheduling tree of the real-time microkernel operating system according to the rate of change of the spatiotemporal tensor residual scalar to reconstruct the instruction flow orchestration, and use the utilization rate of the multi-core computing power core of the operating system and the bus bandwidth occupancy rate as weighting factors to aggregate calculations and generate the core ready queue impedance index that characterizes the degree of blockage of the hardware bus computing power channel.

[0028] Step S3: When the impedance index of the core ready queue is greater than the safety impedance threshold, the kernel scheduling unit triggers the time window decoupled cache fault-tolerant control flow, issues a kernel overwrite instruction to lower the scheduling priority of the kernel scheduling thread corresponding to a spatiotemporal feature fusion method for station environment analysis in the physical bus bandwidth by 3 scheduling levels, and simultaneously injects a step-by-step penalty operator into the computation graph corresponding to the spatiotemporal feature tensor in situ, closes the spatial convolution calculation branch of the secondary sensing feature input graph channel branch, and applies spatial dimension dynamic pruning to the computation graph in the spatial feature extraction step to stabilize the input tensor size within the channel range corresponding to the predetermined basic safety feature subset.

[0029] Preferably, the extraction of the scalar change rate of the spatiotemporal tensor residual in step S1 includes the following sub-steps: Step S11, calculate the tensor difference between the current frame spatiotemporal feature tensor and the feature vector copy of the previous cycle in the feature buffer to obtain the spatiotemporal tensor residual; Step S12, calculate the scalar magnitude of the spatiotemporal tensor residual and divide it by the frame interval time to generate the scalar change rate of the spatiotemporal tensor residual; Step S2, adjusting the weight update of the kernel scheduling tree of the real-time microkernel operating system includes: when the scalar change rate of the spatiotemporal tensor residual increases, proportionally increase the scheduling weight of the kernel scheduling thread in the kernel scheduling tree of the real-time microkernel operating system.

[0030] Preferably, step S3, which involves injecting a stepped penalty operator into the computation graph corresponding to the spatiotemporal feature tensor in situ, includes the following sub-steps: Step S31, obtaining the core task correlation degree corresponding to each secondary sensing feature input graph channel branch; Step S32, classifying each secondary sensing feature input graph channel branch into levels based on the core task correlation degree, and separating the key sensing input graph channel branches and the secondary sensing feature input graph channel branches; Step S33, when the impedance index of the core ready queue is in different overload level ranges, matching and closing the spatial feature extraction of the corresponding number of secondary sensing feature input graph channel branches through the stepped penalty operator.

[0031] Preferably, the cache fault-tolerant control flow that triggers time window decoupling in step S3 includes the following sub-steps: Step S34, when the multi-path spatial feature stream generates non-uniform transmission delay in the physical bus bandwidth transmission, interrupt the time accumulation calculation of the current time window; Step S35, forcibly call the cached feature vector copy of the previous cycle in the feature cache area through a zero-order hold action; Step S36, use the feature vector copy of the previous cycle to replace the multi-dimensional sensing feature data stream that has generated non-uniform transmission delay, and participate in the time accumulation calculation of the current cycle.

[0032] Preferably, after step S3, there is a step to reduce core sensing latency by releasing physical bus bandwidth due to closing the secondary sensing feature input graph channel branch: step S37, centrally schedule the multi-core computing power core of the operating system on the converged basic security feature subset to specifically ensure the deterministic computation latency of the core analysis task.

[0033] Preferably, the kernel scheduling tree of the real-time microkernel operating system adopts a red-black tree structure to organize kernel-mode threads. The weight update of adjusting the kernel scheduling tree of the real-time microkernel operating system in step S2 includes: recalculating the virtual running time of the kernel scheduling thread in the red-black tree structure according to the updated weight, so as to adjust its position in the red-black tree structure.

[0034] Preferably, step S3, which lowers the scheduling priority of the kernel scheduling thread corresponding to a spatiotemporal feature fusion method for station environment analysis in the physical bus bandwidth by 3 scheduling levels, includes: removing the kernel scheduling thread from the real-time scheduling queue of the real-time microkernel operating system and attaching it to the tail of a predetermined low-priority linked list of the time-sharing cyclic scheduling queue.

[0035] Preferably, the safety impedance threshold is a bus crash critical impedance scale determined based on hardware bus dissipation load data under sudden high throughput characteristic flow load conditions in historical periods.

[0036] Preferably, in step S3, converging the input tensor size to a predetermined subset of basic security features includes: after triggering the buffer fault-tolerant control flow, stabilizing the input tensor size within the channel range corresponding to the predetermined subset of basic security features, discarding the spatial convolutional feature tensor channels of the secondary sensing feature input map channel branches, and retaining only the spatial convolutional feature tensor channels containing the core analysis protection region.

[0037] Example 1: In the heterogeneous computing resource scheduling for spatiotemporal feature fusion computing of perceived video streams at urban rail transit stations, under the condition that passengers are suddenly arriving at the station and the doors open, the convergence of features from multiple paths and the accumulation of features from consecutive frames generate nonlinear computing load and bus throughput requirements. The time-sharing scheduling mechanism of the general operating system is independent of the topology of the specific computing model at the upper layer and does not recognize the spatiotemporal coupling characteristics between operators within the computation graph. High-dimensional tensors cause congestion on the data bus and trigger kernel page faults and thread context switching losses. As a result, the time filtering and convergence thread is in a resource starvation state due to resource inversion, causing the task scheduling latency of the spatiotemporal feature fusion pipeline to rise, and the global perception real-time index exceeds the preset safety threshold.

[0038] This method operates on a specific computer system with heterogeneous computing power units. It retrieves the difference data between the spatial feature tensor output by the spatial feature intersection operator and the temporal feature tensor output by the temporal filtering operator within the previous computation cycle through a state acquisition unit, and calculates the rate of change of the spatiotemporal tensor residual scalar. ,in The value is a dimensionless real-time rate of change; the status acquisition unit synchronously reads the task backlog length of the kernel ready queue recorded by the kernel ready queue counter. The kernel scheduling unit receives the rate of change of the spatiotemporal tensor residual scalar. And as the first control variable of the microkernel operating system, it calculates the adaptive activation threshold of the computation graph for the current cycle according to the preset first-order linear calculation rules. In computation graph adaptive activation threshold When the load exceeds the system kernel's preset steady-state threshold, the status acquisition unit retrieves the system data bus bandwidth utilization percentage from the heterogeneous computing unit, and the kernel scheduling unit adjusts the task backlog length. As a fundamental hardware blocking parameter, a multiplicative nonlinear weighted calculation is performed in conjunction with the system data bus bandwidth utilization percentage to output the core ready queue impedance index. ,in Impedance quantization rating scalar; impedance index in the core ready queue When the impedance exceeds the safety threshold, the kernel scheduling unit triggers the cache fault tolerance control flow. This control flow is an asynchronous exception handling thread residing in the operating system kernel scheduling layer. It relies on the real-time data arrival timestamps of multiple feature channels monitored by a high-precision timer in the underlying hardware as feature input parameters. When it determines that the absolute value of the data arrival time difference between any two feature channels is greater than the preset discrete transmission delay tolerance threshold of 4ms, it immediately activates the timing blocking mechanism. The specific processing actions include forcibly interrupting the cumulative summation operation instruction flow within the current hardware time window, and simultaneously directly calling the pre-allocated historical feature buffer protected by contiguous physical memory pages through a high-priority kernel interrupt. The previous cycle's feature vector copy is used to completely replace the current cycle's multidimensional sensing feature data stream, which experiences non-uniform delays, in subsequent time feature accumulation operator operations. This cuts off data waiting and blocking and scheduling lock contention in software timing, resulting in a stable feature output in a zero-order hold state. Subsequently, the kernel scheduling unit issues a kernel overwrite instruction to lower the scheduling priority of the kernel scheduling thread corresponding to the spatiotemporal feature fusion method by three levels. The kernel scheduling thread is then removed from the microkernel operating system's real-time scheduling queue and attached to the tail of the low-priority linked list of the time-sharing cyclic scheduling queue, reducing the bus occupation of high-traffic spatial feature computation tasks. Finally, the kernel scheduling thread is attached to the time-sharing cyclic scheduling queue. While the actions at the tail of the low-priority linked list in the queue macroscopically reduce the frequency of processor preemption by this thread, the synchronous injection of a step-by-step penalty operator into the computation graph forces the feature parameters of secondary perceptual feature input graph channel branches that need to be closed to be set to zero during forward propagation. Because a large amount of data flow undergoes dynamic spatial dimensional pruning in physical memory, the computational load and data bus throughput requirements of this thread each time it is awakened in the time-sharing queue are reduced by more than 50%, and the physical execution time of a single forward propagation operation is drastically reduced from 15 milliseconds to 0.8 milliseconds. The microkernel operating system centrally schedules the released hardware computing power to another independent hard real-time core. The protection thread is specifically designed to receive and process a condensed subset of basic security features. This collaborative mechanism, which uses significant dimensionality reduction pruning at the algorithmic level to offset the decrease in the scheduling level of the underlying operating system, fundamentally eliminates the risk of task starvation that may be caused by threads being mounted to time-sharing queues. This ensures that the global awareness scheduling latency jitter of the core security detection task remains stably controlled within the system design safety benchmark of less than 2 milliseconds. The real-time microkernel operating system kernel scheduling architecture includes a two-layer collaborative scheduling structure composed of a hard real-time preemptive queue and a kernel scheduling tree. The kernel scheduling thread corresponding to the spatiotemporal feature fusion method resides in the hard real-time preemptive queue when the hardware is not overloaded, when the rate of change of the spatiotemporal tensor residual scalar... When the preset steady-state discrimination threshold is exceeded, the kernel scheduling unit schedules the kernel scheduling thread to a dynamic smooth scheduler using a red-black tree structure and calculates the virtual runtime. Adjusting the dynamic position of kernel scheduling threads in the red-black tree structure, and virtual runtime. Satisfy the following formula: ,in, This is the updated virtual runtime. This is the original virtual runtime for the current period. Set the base runtime increment for the dynamic smooth scheduler. The preset dynamic step size adjustment coefficient is a real number ranging from 0.5 to 1.5. To increase the scalar change rate of the spatiotemporal tensor residual, the virtual runtime is increased to cause the kernel scheduling thread to deflect to the right node in the red-black tree structure to reconstruct the instruction flow orchestration. This proportionally increases the scheduling weight of the kernel scheduling thread in the kernel scheduling tree of the real-time microkernel operating system. In essence, this means increasing the punitive control weight of the kernel scheduling thread in the overall system resource optimization configuration, rather than increasing its preemptive priority for acquiring processor execution rights in the traditional sense. Under the completely fair scheduling logic of the microkernel operating system, this is achieved by manually increasing... The virtual execution time of this thread is rapidly accumulated, causing it to quickly shift to a lower priority execution node on the right side of the red-black tree topology. This physically shortens the time it occupies the central processing unit in a single time-sharing session. The core purpose of this increased punitive weight is to free up valuable bus bandwidth and computing units for core analysis tasks by actively lengthening the response interval of this non-core thread. This achieves hardware-software impedance matching between the topological causal characteristics of the algorithm layer and the physical resource allocation of the operating system. In the underlying control implementation of the red-black tree dynamic smooth scheduler, this is specifically manifested by increasing the dynamic step size adjustment coefficient. To amplify the rate of change of the spacetime tensor residual scalar Virtual runtime The penalized contribution weight accelerates the accumulation of virtual runtime of the kernel scheduling thread corresponding to the feature fusion method under high load and deflects it to the right node of the red-black tree structure. This reduces its running priority and restructures the instruction flow arrangement at the operating system kernel level, thereby forcibly releasing bus computing resources to prioritize the deterministic progress of core analysis tasks. At the same time, a step-by-step penalty operator is injected in situ into the computation graph corresponding to the spatial feature tensor to calculate the core task correlation of each perceptual feature input graph channel branch. Based on the core task correlation, a specific number of perceptual feature input graph channel branches are closed and their spatial feature extraction is stopped. Dynamic spatial dimensional pruning is performed on the computation graph to adjust the input tensor size to the channel range corresponding to the basic security feature subset, releasing physical bus bandwidth. The topology and dependencies of the specific computation model are translated into the constraints of the system kernel-level variable scheduling tree, completing the microkernel operating system resource scheduling tree and... Hardware and software impedance matching for causal topology of specific computational models; a control flow interface is established between the kernel space of the real-time microkernel operating system and the application-layer deep learning inference engine through a pre-registered dedicated shared memory page. When the kernel scheduling unit lowers the running priority of the kernel scheduling thread corresponding to the feature fusion method at the kernel layer, the kernel scheduling unit atomically writes an instruction reassembly status word to a specific control pointer position of the dedicated shared memory page; before each execution of forward propagation operator flow orchestration, the application-layer deep learning inference engine reads the status word through a lightweight system call. Once a change in the status word is detected, the current general instruction pipeline is immediately suspended, and a preset task reassembly operator is called to switch the original serial concurrent multiple spatial convolution instruction flows to a flexible reassembly instruction flow with conditional branch prediction, thereby completing the closed-loop mapping from the adjustment of the underlying thread priority to the reconstruction of the upper-layer algorithm instruction flow at the application layer.

[0039] Through the core ready queue impedance index The triggered pipeline adjustment and adaptive scheduling dynamic interaction, the linkage response of passenger flow space pressure to kernel state feature cache calls, thread priority adjustment and computation graph dynamic pruning, and the construction of continuous physical memory pages enable the spatial feature tensors of the specified analysis protection area to be transmitted via hardware direct memory access, reducing the transient response lag of the kernel scheduling unit under high concurrency and the task stoppage caused by bus overload. By utilizing the existing heterogeneous computing power units and status acquisition units in the computer system, the global perception task scheduling latency jitter under sudden working conditions is controlled within 2ms, reducing the scheduling latency jitter of the system under sudden load impact and maintaining the reliability of the station environment analysis method. In actual operation, in order to accurately determine the occurrence of non-uniform transmission latency, the status acquisition unit uses the underlying hardware high-precision timer to cyclically monitor the real-time data arrival timestamps of multiple spatial feature streams on the physical bus. When the absolute value of the data arrival time difference between any two feature channels is calculated to be greater than the preset transmission latency tolerance threshold of 4ms, the system determines that non-uniform transmission latency has occurred and immediately blocks the time accumulation calculation of the current time window, forcibly calling the previous cycle feature copy to perform zero-order hold action.

[0040] Example 2: When the system is in a heterogeneous computing resource scheduling test condition for high-dimensional spatiotemporal feature fusion computing of multi-source sensing video streams in urban rail transit stations, the test platform is built on a rack-mounted multi-core processor and a graphics acceleration chip. The processor core includes 16 independent computing cores with a core base frequency of 2.4 GHz and a system physical bus bandwidth of 64 GB per second, corresponding to the concurrent high-flow passenger feature convergence environment when trains enter the station during peak hours and multiple passage doors open. The original input data used in the experiment comes from a publicly available rail transit passenger flow sensing dataset. Gaussian background noise with a signal-to-noise ratio of 20 dB is superimposed on the original feature data stream, and simulated power frequency interference harmonics with a frequency of 50 Hz are injected to create a non-ideal input environment. In the test parameter configuration, the sampling period of the state acquisition unit is... As a process control parameter, sampling period The sampling period is determined based on the nonlinear relationship between the system bus bandwidth utilization margin and the data flow update intensity. When the system bus bandwidth utilization margin decreases or the data flow update intensity increases, the sampling period... To improve scheduling response speed, the sampling period is reduced when the bus bandwidth utilization margin increases and the data stream update intensity decreases. Increase the sampling period to reduce computing power overhead; under the current concurrent operating conditions, based on the corresponding relationship, the sampling period... The value was set to 10ms, which was used as the input control parameter for the test platform.

[0041] Under the same non-ideal input environment, the test was divided into a control group and the present invention sample group. In the control group, when concurrent throughput occurred on the data bus, the processor was unable to obtain the spatiotemporal topological causal relationships of the computational model, resulting in high-dimensional tensor accumulation causing bus overload. The measured feature data signal-to-noise ratio decreased to 12.3dB, and its task scheduling latency jitter was 18.42ms. The present invention sample group was divided into three different concurrent ready gradient states. In sample group one, the task backlog length of the core ready queue was... There are 5, and the rate of change of the spatiotemporal tensor residual scalar is extracted by the state acquisition unit. The threshold is 0.12, representing the adaptive activation threshold of the computation graph calculated by the kernel scheduling unit according to the first-order linear computation rule. The calculated impedance index for the core ready queue is 0.20. The value is 15.6. When the former is less than the impedance safety threshold of 70.0, the system is in a non-overload state, and no adjustment is made to the computation graph structure. The measured task scheduling delay jitter is 0.32ms. In sample group two of this invention, the task stacking length is... The number was changed to 25, and the rate of change of the extracted spatiotemporal tensor residual scalar was... The threshold was changed to 0.54, and the resulting computational graph adaptive activation threshold was calculated. With a value of 0.55, when the former is greater than the steady-state discrimination threshold of 0.50, the percentage of system data bus bandwidth utilization is 76.4%, and the calculated core ready queue impedance index is... The impedance is 74.2; in the core ready queue impedance index. Under conditions where the impedance safety threshold of 70.0 is greater than the threshold value, the kernel scheduling unit reduces the thread scheduling priority by 3 levels and changes the thread from the real-time scheduling queue to the time-sharing cyclic scheduling queue. Simultaneously, it writes a stepped penalty operator to the computation graph in situ and disables spatial feature extraction for 3 secondary feature channels. The measured feature data signal-to-noise ratio becomes 18.7 dB, and the task scheduling latency jitter is 1.15 ms. In sample group three of this invention, the task backlog length... To protect the upper limit of 45, the scalar rate of change of the spatiotemporal tensor residuals. The calculated adaptive activation threshold for the computational graph is 0.88. The value is 0.85, the system data bus bandwidth utilization percentage is 91.8%, and the weighted calculation yields the core ready queue impedance index. The impedance is 116.5; in the core ready queue impedance index. Under the condition that the impedance safety threshold of 70.0 is greater than the threshold, the kernel scheduling unit lowers the thread scheduling priority, the system shuts down the spatial feature extraction of 6 secondary feature channels, the input tensor size is reduced to a subset of the basic safety features, and the task scheduling latency jitter is 1.88ms.

[0042] An out-of-range control group was added to the test, when the task accumulation length in the working condition was [not specified]. The number of errors exceeded the design limit by 60, and the rate of change of the spatiotemporal tensor residual scalar was also affected. When it reaches 0.96, the calculated core ready queue impedance index The value was 158.4; the system generated a nonlinear overload saturation state, the spatial dimension dynamic pruning adjustment of the stepped penalty operator increased, the input tensor of the main feature channel was interrupted, the system kernel generated buffer data overflow and abnormal page interrupt, and the measured task scheduling delay jitter was 14.65ms; the data of the out-of-range control group showed that the system's operational stability deteriorated after the key parameters exceeded the upper limit of the protection range, while the sample groups of this invention 1, sample groups 2 and sample groups 3 of this invention within the protection range, through the impedance matching of the instruction flow of the microkernel operating system and the computing topology, in the non-ideal operating environment accompanied by background noise and simulated power frequency interference harmonics, the data diversion of the heterogeneous computing bus and the pruning operator formed a causal relationship, maintaining the order of feature computing flow instruction arrangement during the throughput gradient change process, and the computer system recovered to a steady-state ready state without queue backlog at the end of the test condition.

[0043] Example 3: Under the condition that the safety impedance threshold is selected from 60.0 to 80.0 and the steady-state discrimination threshold is selected as 0.50, the system computing core calculates multiple sets of baseline characteristic fusion calculation data streams with load changes ranging from 10% to 100% during the power-on self-test period. The bus bandwidth utilization percentage and the task backlog length of the core ready queue are collected by the kernel ready queue counter and the bus performance monitoring register at a sampling frequency of 10Hz. Obtain bus transmission delay as a function of task backlog length The characteristics of the increase, where the bus bandwidth utilization percentage reaches 85% and the task backlog length... When the number of tasks reaches 25, the bus transmission delay increases with the length of the task backlog. The sudden increase in the first derivative of the rate of change indicates that the hardware computing power channel has reached the overload boundary. The system obtains a weighted accounting value based on the bus bandwidth utilization percentage of 76.4% at the overload boundary and the task backlog length of 25. This weighted accounting value corresponds to a safe impedance threshold of 70.0, and the rate of change of the spatiotemporal tensor residual scalar at the overload boundary is also considered. Corresponding to a steady-state discrimination threshold of 0.50, the logical criterion value for real-time scheduling is established, and the upper-level specific calculation model includes... The computational graph of a neural network in which a two-dimensional convolutional operator in layer space and a long-term memory operator in time are orthogonally cascaded. The neural network has 8 independent video camera channels located in the platform's protected area. Each channel outputs a 512×7×7 third-order floating-point tensor as its image spatial feature tensor in the current cycle. The real-time microkernel operating system has a specific scheduling tree structure residing in its kernel mode. This scheduling tree structure has a pointer interface for adjusting weights based on the algorithm's causal topology. The system also monitors the task backlog length in the core ready queue, which is read in real-time by the status acquisition unit. Increased to 35 and calculated core ready queue impedance metrics When the threshold reaches 92.4 and exceeds the safe impedance threshold of 70.0, this threshold is used as a static critical control boundary value stored in the kernel global configuration register. Its objective calibration relies on a benchmark database established by continuously reading the bus bandwidth utilization and task queue queuing status for 100 cycles at a sampling frequency of 10Hz by the bus performance monitoring register during system power-on self-test or no-load ready state. When the system detects a hardware overload condition where the bus bandwidth utilization percentage reaches 85% and the kernel core ready queue task backlog length reaches 25, the first derivative of the rate of change of bus transmission delay with the increase of task backlog length will produce a step increase. At this time, the system will... The bus bandwidth utilization percentage of 76.4% at the aforementioned boundary point and the task backlog length of 25 are incorporated into the aforementioned multiplicative nonlinear weighted calculation. A quantized scale value of 70.0, representing the hardware bus computing power channel reaching the collapse boundary, is then determined. This value serves as a hard control breakpoint. When the core ready queue impedance index crosses this breakpoint, the closed-loop system is triggered to enter overload degradation protection mode. At this time, the microkernel operating system starts the spatial dimension dynamic pruning scheduler. The specific steps include: First, the kernel scheduling unit calls the target object pixel change frequency stored in the history buffer register for each video camera channel within the past 50 consecutive sampling periods, and calculates the core task correlation degree corresponding to each channel. The value ranges from 0 to 1; specifically, the frequency of pixel changes of the target object is statistically analyzed in real time by the application layer image inference engine residing in user space during forward propagation calculation, and asynchronously synchronized to the historical buffer register mapping area of ​​the kernel state via the cross-layer dedicated shared memory page reserved by the microkernel operating system during the initialization phase. This allows the kernel scheduling unit in kernel space to securely retrieve cross-layer business state data by directly addressing this kernel state image area. Under the premise of strictly maintaining the physical space security isolation between kernel state and user state, a seamless transmission path from the algorithm layer topology state to the system's underlying scheduling tree is established. In the second step, the kernel scheduling unit will determine the correlation of core tasks. The eight video camera channels are sorted in ascending order of numerical value, generating a descending sorted linked list containing the level order; the kernel scheduling unit then uses the current kernel ready queue impedance index... The absolute value of the difference from the safe impedance threshold of 70.0 is used to determine the number of physical channels to be cut off, which is 4, based on a preset discrete step function. In the fourth step, the kernel scheduling unit sequentially extracts the first 4 low-association video camera channels corresponding to the degraded sorting list and injects a step-by-step penalty operator into the corresponding neural network computation graph. This operator is essentially a memory addressing control module pre-residing in the forward propagation computation graph of the application layer image inference engine. It establishes control flow by reading the control pointer synchronized across kernel space to the dedicated shared memory image area. Its bound objective feature input parameters are the specific discrete difference range of the core ready queue impedance index deviating from the safe impedance threshold and the corresponding video camera channel degraded sorting list index. The specific processing action is as follows: The operating system's memory management unit page table remapping mechanism is invoked, and based on the step size of the difference range, the first 4 secondary sensing feature input graph channels in the degraded sorting list are selected. The physical starting address of the spatial two-dimensional convolution operator corresponding to the branch in physical memory is redirected and modified. Specifically, the addressing stride and control mask of these secondary channels during forward propagation computation are atomically overwritten, thereby forcibly redirecting their addressing target to the system-preset zero-constant aligned physical memory page. The input feature tensor data of the corresponding channel is directly forced to zero during the hardware read phase, thereby closing the spatial convolution computation of the secondary branch and closing the convergence state where all output feature parameters are cleared. This step-by-step penalty operator adjusts the physical starting address of the aforementioned spatial two-dimensional convolution operator in physical memory. The kernel scheduling unit injects the step-by-step penalty operator physical path into the computation graph corresponding to the spatiotemporal feature tensor in situ, using the memory management unit page table remapping mechanism. During the computer system initialization phase, the application layer inference engine registers a dedicated shared memory page and reserved operator control pointer with the real-time microkernel operating system through system calls. When the core ready queue impedance index When the safe impedance threshold is exceeded, the kernel scheduling unit modifies the physical starting address of the spatial 2D convolution operator in physical memory, converts the difference between the two into a control mask, and writes it into the reserved operator control pointer. The control mask is used to adjust the physical addressing stride of the spatial 2D convolution operator. This causes the application layer inference engine to redirect the physical memory address of the secondary perceptual feature input map channel branch that needs to be closed to a system-preset zero-constant aligned page when performing forward propagation calculations. The corresponding input feature tensor parameters are forcibly set to zero to disable spatial feature extraction. In actual operation, to avoid large-scale TLB invalidation and anomalies caused by frequent page table modifications, [further measures are taken]. The system overload dissipation caused by page interrupts is mitigated by the memory management unit's page table remapping mechanism, which is triggered only once during the state transition when the kernel ready queue impedance index crosses the safe impedance threshold, rather than being repeatedly called frequently as the data frame cycle progresses. Furthermore, the microkernel operating system pre-establishes shadow page table entries during the initialization phase, requiring only an atomic change to the global pointer of the top-level page table directory when a state transition is triggered. This keeps the context switching overhead of the overall remapping operation at the microsecond level, thereby physically eliminating memory management dissipation and ensuring that the task scheduling latency jitter of the sensing thread is stably controlled within two milliseconds. Within the design safety baseline, the input feature tensor parameters corresponding to the four low-association video camera channels are set to zero, and the spatial feature extraction stream of the corresponding camera branches is shut down. To establish a deterministic mapping between the video camera channel branches in the high-level algorithm topology and the page table entries of the underlying hardware memory management unit, a two-layer structure mapping table is constructed during the initialization phase. The first layer of this mapping table records the virtual address logical block codes of the spatial two-dimensional convolution operators corresponding to each of the eight video camera channels during forward propagation computation. The second layer directly binds these virtual addresses to the physical memory management unit through kernel-mode memory descriptors. The mapped top-level page table directory entry pointer; when the kernel scheduling unit issues an instruction to modify the control mask, it can directly address the specific top-level page table corresponding to the first four low-association video camera channels through this mapping table, realizing accurate cross-layer data structure binding and in-situ coverage; in the fifth step, the kernel scheduling unit limits the input tensor scale to the basic security feature subset composed of the four core security detection camera channels, reduces the data transmission task flow scale of the physical bus by 50.0%, blocks the convergence of overload traffic spatial features to the data bus, and completes the reverse closed-loop constraint of hardware bearing resistance on the operator computation dimension.

[0044] The kernel scheduling unit collects the current system data bus bandwidth utilization percentage of the heterogeneous computing units. Meanwhile, in order to incorporate the computing power contention and dissipation caused by concurrent execution of multiple cores within the operating system into the blocking assessment system, the status acquisition unit synchronously reads the utilization rate of the multi-core computing power cores in the current cycle through the kernel-level performance monitoring register of the microkernel operating system. The kernel scheduling unit will optimize the utilization of multi-core computing power. Percentage of system data bus bandwidth utilization The multiple independent variables that together serve as multiplicative nonlinear weighting factors, and the task stack length that serves as a basic term for hardware blocking. A scalar convergence product calculation is performed to generate an integer quantized rating value for assessing hardware channel congestion, which runs within the kernel scheduling unit. The calculation logic for this value is to use the current ready queue task backlog length as a base factor, multiplied by a combined weighted gain term constructed from the multi-core computing power utilization component and the bus bandwidth utilization component. Specifically, the value of this combined weighted gain term is equal to 1.0 plus the product of the multi-core resource mutual exclusion calibration coefficient and the multi-core computing power utilization read from the kernel-level performance monitoring register, plus the product of the massive physical bus congestion calibration coefficient and the current cycle system data bus bandwidth utilization percentage. Wherein, the kernel core ready queue... The task stack length is a discrete integer between 0 and 60. The multi-core utilization and bus bandwidth utilization are both floating-point numbers between 0.0 and 1.0. The multi-core resource mutual exclusion calibration coefficient ranges from 1.0 to 2.0, and the physical bus congestion calibration coefficient ranges from 2.0 to 3.0. These two calibration coefficients are dynamically corrected during the computer system initialization self-test period by continuously reading the average of idle parameters over 100 sampling periods as a bias. This ensures that the final aggregated integer quantization rating value remains stable within a discrete distribution range of 0.0 to 160.0. This value serves as the sole criterion for switching the operating system's underlying control flow, and its value varies with the multi-core utilization. Increase and system data bus bandwidth utilization percentage The increase is monotonically increasing, thereby quantifying in situ the combined blocking effect caused by multi-core resource mutual exclusion and physical bus congestion, resulting in the final output core ready queue impedance index. The impedance safety threshold determination process has complete resource status support. In practical engineering applications, the value boundaries of the multi-core resource mutual exclusion calibration coefficient and the physical bus congestion calibration coefficient have clear engineering failure and side effect limitation significance. The lower bound of the multi-core resource mutual exclusion calibration coefficient is set to 1.0. If it is lower than 1.0, the system will not be able to effectively quantify the mutual exclusion lock overhead generated when multiple computing cores execute concurrently, resulting in a serious underestimation of the impedance index under heavy load conditions, thus missing the best pruning opportunity. Its upper bound is set to 2.0. If it is higher than 2.0, it will excessively amplify the normal thread switching loss, causing the system to generate frequent false triggers when it is not overloaded. The lower bound of the physical bus congestion calibration coefficient is set to 2.0 to ensure the minimum sensitivity to transient high throughput characteristic flow. If it is lower than 2.0, it will not be able to intercept the bus overload in time. Its upper bound is set to 3.0. If it is higher than 3.0, it will cause the integer quantization rating value to easily exceed the theoretical upper limit of 160.0 and cause numerical overflow, resulting in frequent control flow deadlock.

[0045] The adaptive pruning control of the neural network computation graph is transformed into one based on the core task relevance. The system modifies the physical memory addressing base address, responds to bus overload during scheduling, intercepts low-correlation channel feature data streams and completes dimensionality reduction and convergence, and uses a microkernel operating system kernel to adjust a step-by-step penalty operator to block disordered thread context switching overhead, thereby offsetting the overload dissipation of physical bus bandwidth. This allows the instruction stream of the core security detection task to obtain continuous computing core priority scheduling rights, transforms the hardware-level blocking state of the heterogeneous computing bus in concurrent scenarios into a predictable gradient degradation steady-state response, and keeps the predetermined computation latency of the perception thread within the system design safety benchmark of less than 2ms.

[0046] Example 4: When the system faces on-site deployment and benchmark parameter calibration in a completely new station environment, due to hardware limitations in the number of security cameras mounted on different stations and the physical transmission capacity of the data bus, the status acquisition unit continuously reads the bus bandwidth utilization and task queue queuing status for 100 cycles in the system's idle and ready state, calculates the average idle bus impedance, and uses the average idle bus impedance as a bias to correct the core ready queue impedance index. The dynamic calculation parameters ensure that the adaptive activation threshold determination logic conforms to the physical load-bearing capacity of the current hardware. When the system is under long-term operating conditions, the aging of physical components and low-frequency drift of environmental noise interfere with the control state. The kernel scheduling unit opens a 24-hour time-series sliding window, and the status acquisition unit reads the task backlog length of the core ready queue during off-peak hours. The kernel scheduling unit calculates the real-time environmental background impedance using a preset time decay factor. Periodically reconstruct the safety impedance threshold, correct the system operating baseline to the current physical state, and keep the sensing thread scheduling latency within 2ms under high-concurrency passenger flow conditions.

[0047] Example 5: When the system faces an unsteady load condition caused by the long-term operation of multiple sensing nodes at urban transportation stations, resulting in temperature drift in the underlying hardware characteristics and increased conflicts between the convergence of high-dimensional spatiotemporal features of multiple paths, the lens of the security video camera channel experiences temperature drift in the characteristic voltage output of the image sensor due to temperature changes. This causes the extracted spatial feature tensor to contain spatiotemporal variation errors accumulated with consecutive frames. Since the general operating system does not have a channel to directly obtain the spatiotemporal multidimensional data features of the computational model, when the data bus generates high concurrency throughput and faces multi-core concurrent lock contention in the feature fusion pipeline, the processor allocation unit experiences increased task scheduling latency, causing the spatiotemporal feature fusion pipeline to fall into a core resource preemption state.

[0048] In the parameter adaptation and initialization phase of new application scenarios, the system constructs an adaptive parameter matrix to establish a running benchmark. After the interaction between the spatial feature tensor and the temporal feature tensor of adjacent calculation cycles, a discrete difference distribution is generated. The system introduces the spatiotemporal feature convergence entropy. To quantify the variability of the discrete difference distribution, the state acquisition unit reads the spatial feature tensors of the four video camera channels for the current period and retrieves copies of the historical feature tensors from the previous period. In the processor's computational core, the variance of the difference at corresponding element positions is calculated. When the measured variances of the differences for the four video camera channels are 0.15, 0.12, 0.18, and 0.11, respectively, the data processing unit performs a standardized root mean square concatenated summation operation on these four independent variance values. Specifically, the known engineering processing steps are as follows: first, the variance of the difference for each channel is squared to eliminate negative fluctuations, resulting in four squared positive components. Then, these four positive components are summed to obtain a total value. This total value is then divided by the total number of channels (4) to obtain the mean square term. Finally, the arithmetic square root of this mean square term is extracted, thus completing the normalized scalar convergence of the multi-channel feature dispersion and directly outputting the spatiotemporal feature intersection entropy. The dimensionless value is 0.45; the kernel scheduling unit receives the spatiotemporal feature intersection entropy. And as the third control variable of the operating system, combined with the aforementioned spatiotemporal tensor residual scalar rate of change. With the length of the task backlog in the ready queue The weight update matrix of the operating system's kernel scheduling tree is corrected in situ according to the dimensional homogeneous balance criterion, when the task backlog length of the ready queue is... For 25 and the scalar change rate of the spatiotemporal tensor residuals When the value is 0.54, the kernel scheduling unit will set the kernel ready queue impedance index. The real-time update is 74.2; the weight update matrix of the operating system's kernel scheduling tree is a four-dimensional mapping control matrix initialized in kernel mode, consisting of four independent control dimensions: the task backlog length component of the ready queue, the rate of change component of the spatiotemporal tensor residual scalar, the spatiotemporal feature intersection entropy component, and the dynamic step size adjustment coefficient component. According to the dimensional homogeneity balance criterion, this weight update matrix has a pre-set set of normalized adjustment step size parameters. When the spatiotemporal feature intersection entropy increases from 0.10 to 0.45, the kernel scheduling unit directly addresses this update matrix, dynamically increasing the multiplicative adjustment factor linked to the virtual runtime penalty term from the base 1.0 to 2.5. This increase directly affects the step size accumulator of the red-black tree dynamic smoothing scheduler, causing a sharp increase in the slope of the virtual runtime accumulation of the non-core feature fusion thread in the red-black tree, thus achieving precise in-situ correction of the underlying operating system matrix scheduling parameters using the complex feature entropy value of the algorithm layer. Due to the core ready queue impedance index... When the impedance exceeds the safe impedance threshold of 70.0, the microkernel operating system lowers the priority of the thread corresponding to the spatiotemporal feature fusion method by three levels and moves it to the tail of the low-priority linked list of the time-sharing cyclic scheduling queue. Simultaneously, it dynamically adjusts the physical addressing stride of the spatial two-dimensional convolution operator using a step-by-step penalty operator. It also shuts down the first four video camera channels in the current degraded sorting list, enabling high-volume data streams to undergo dynamic spatial dimensional pruning in physical memory to release bus bandwidth. The spatiotemporal feature intersection entropy, The aforementioned rate of change of the spatiotemporal tensor residual scalar. The length of the task backlog in the kernel's ready queue. The impedance index value for the core ready queue.

[0049] Due to the convergence of entropy through spatiotemporal features Hardware channel impedance matching is achieved between the system kernel layer scheduling weights and the dynamic feature pruning channel. Even under operating conditions where environmental temperature interference causes time-varying degradation of the characteristics of the underlying hardware devices, the causal topology of the specific computation model and the physical resource scheduling of the real-time microkernel operating system still maintain a stable correlation. The direct memory access mechanism and the contiguous physical memory pages where the video camera channel feature copies are located generate hardware and software collaboration. By utilizing the existing processor resources in the system, the overload occupation of the data bus bandwidth by high-concurrency feature computation flow is eliminated, the throughput performance of heterogeneous computing units is improved, and the task scheduling latency jitter in the passenger flow feature analysis process is kept within a safe index window of less than 2ms. The system core ready queue is restored to a smoothly reduced steady-state ready state at the end of the test.

[0050] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A spatiotemporal feature fusion method for platform environment analysis, characterized in that, Includes the following steps: Step S1: Obtain the spatiotemporal feature tensor generated by the intersection of the spatial and temporal dimensions of the multidimensional sensing feature data stream corresponding to the platform environment input channel, and extract the rate of change of the spatiotemporal tensor residual scalar generated by the accumulation of the spatiotemporal feature tensor with consecutive frames during runtime. Step S2: Adjust the weight update of the kernel scheduling tree of the real-time microkernel operating system according to the rate of change of the spatiotemporal tensor residual scalar to reconstruct the instruction flow orchestration, and use the utilization rate of the multi-core computing power core of the operating system and the bus bandwidth occupancy rate as weighting factors to aggregate calculations and generate the core ready queue impedance index that characterizes the degree of blockage of the hardware bus computing power channel. Step S3: When the impedance index of the core ready queue is greater than the safety impedance threshold, the kernel scheduling unit triggers the time window decoupled cache fault-tolerant control flow, issues a kernel overwrite instruction to lower the scheduling priority of the kernel scheduling thread corresponding to a spatiotemporal feature fusion method for station environment analysis in the physical bus bandwidth by 3 scheduling levels, and simultaneously injects a step-by-step penalty operator into the computation graph corresponding to the spatiotemporal feature tensor in situ, closes the spatial convolution calculation branch of the secondary sensing feature input graph channel branch, and applies spatial dimension dynamic pruning to the computation graph in the spatial feature extraction step to stabilize the input tensor size within the channel range corresponding to the predetermined basic safety feature subset.

2. The spatiotemporal feature fusion method for platform environment analysis according to claim 1, characterized in that, The extraction of the scalar change rate of the spatiotemporal tensor residual in step S1 includes the following sub-steps: Step S11, calculate the tensor difference between the current frame spatiotemporal feature tensor and the feature vector copy of the previous period in the feature buffer to obtain the spatiotemporal tensor residual; Step S12, calculate the scalar magnitude of the spatiotemporal tensor residual and divide it by the frame interval time to generate the scalar change rate of the spatiotemporal tensor residual. The weight update of the kernel scheduling tree of the real-time microkernel operating system in step S2 includes: when the rate of change of the spacetime tensor residual scalar increases, the scheduling weight of the kernel scheduling thread in the kernel scheduling tree of the real-time microkernel operating system is increased proportionally.

3. The spatiotemporal feature fusion method for platform environment analysis according to claim 1, characterized in that, Step S3, which involves injecting a stepped penalty operator into the computation graph corresponding to the spatiotemporal feature tensor in situ, includes the following sub-steps: Step S31, obtaining the core task correlation degree corresponding to each secondary sensing feature input graph channel branch; Step S32, classifying each secondary sensing feature input graph channel branch into levels based on the core task correlation degree, and separating the key sensing input channel branch and the secondary sensing feature input graph channel branch; Step S33, when the impedance index of the core ready queue is in different overload level ranges, matching and closing the spatial feature extraction of the corresponding number of secondary sensing feature input graph channel branches through the stepped penalty operator.

4. The spatiotemporal feature fusion method for platform environment analysis according to claim 2, characterized in that, The cache fault-tolerant control flow that triggers time window decoupling in step S3 includes the following sub-steps: Step S34, when the multi-path spatial feature stream generates non-uniform transmission delay in the physical bus bandwidth transmission, the time accumulation calculation of the current time window is interrupted; Step S35, the previous cycle feature vector copy cached in the feature buffer is forcibly invoked through a zero-order hold action; Step S36, the previous cycle feature vector copy is used to replace the multi-dimensional sensing feature data stream that has generated non-uniform transmission delay and participates in the time accumulation calculation of the current cycle.

5. The spatiotemporal feature fusion method for platform environment analysis according to claim 1, characterized in that, After step S3, the process includes steps to reduce core sensing latency by releasing physical bus bandwidth due to closing the secondary sensing feature input graph channel branch: Step S37, centrally schedule the multi-core computing power cores of the operating system on the converged basic security feature subset to specifically ensure the deterministic computation latency of the core analysis task.

6. The spatiotemporal feature fusion method for platform environment analysis according to claim 1, characterized in that, The kernel scheduling tree of the real-time microkernel operating system uses a red-black tree structure to organize kernel-mode threads. Step S2, adjusting the weight update of the kernel scheduling tree of the real-time microkernel operating system, includes: recalculating the virtual running time of the kernel scheduling thread in the red-black tree structure according to the updated weight, so as to adjust its position in the red-black tree structure.

7. The spatiotemporal feature fusion method for platform environment analysis according to claim 1, characterized in that, Step S3 involves lowering the scheduling priority of the kernel scheduling thread corresponding to the spatiotemporal feature fusion method used for station environment analysis in the physical bus bandwidth by 3 scheduling levels. This includes removing the kernel scheduling thread from the real-time scheduling queue of the real-time microkernel operating system and attaching it to the tail of the predetermined low-priority linked list of the time-sharing cyclic scheduling queue.

8. The spatiotemporal feature fusion method for platform environment analysis according to claim 1, characterized in that, The safety impedance threshold is a bus crash critical impedance scale determined based on hardware bus dissipation load data under sudden high throughput characteristic flow load conditions during historical periods.

9. The spatiotemporal feature fusion method for platform environment analysis according to claim 1, characterized in that, Step S3, which involves converging the input tensor size to a predetermined subset of basic security features, includes: after triggering the buffered fault-tolerant control flow, stabilizing the input tensor size within the channel range corresponding to the predetermined subset of basic security features, discarding the spatial convolutional feature tensor channels of the secondary sensing feature input map channel branches, and retaining only the spatial convolutional feature tensor channels containing the core analysis protection region.

Images