An acoustic-optical combined bird repelling method and system based on size model cooperation and a medium

Abstract

The application discloses a kind of based on size model cooperation's sound-light combined bird repelling method, system and medium, it is related to intelligent bird repelling technical field, including: control cloud platform camera cruise collection image;Edge computing module runs front-end small model to image real-time detection, when detecting first confidence is greater than or equal to first threshold, it is determined as suspected target;Suspected target image region is uploaded to rear-end server;Rear-end server runs back-end large model to the image received is reviewed, if second confidence after review is greater than or equal to second threshold, then confirm as bird target;If second confidence is greater than the first threshold and less than second threshold, then combine time sequence multi-frame information to final determination;After confirming bird target, according to bird class executes differentiating sound-light repelling operation;Rear-end server is based on review result and carries out asynchronous incremental training, and updates front-end small model by knowledge distillation;The method improves bird repelling effect and equipment life.

Description

Technical Field

[0001] This invention relates to the field of intelligent bird deterrence technology, and in particular to a method, system and medium for combined sound and light bird deterrence based on the collaboration of large and small models. Background Technology

[0002] As the central hub of the power system, the safe and stable operation of substations is of paramount importance. However, substations have long been plagued by bird damage: birds build nests on equipment supports, and nesting materials such as dead branches and wires can easily cause short circuits; bird droppings accumulate on the surface of insulators, reducing insulation performance and causing flashover tripping; birds accidentally colliding with high-voltage equipment may directly cause power outages.

[0003] Substations are densely packed with equipment and operate at high voltage levels. Bird-related short circuits or tripping incidents can cause widespread power outages and significant economic losses. Therefore, developing efficient and reliable bird-repelling systems specifically designed for the unique environment of substations is of great practical importance.

[0004] Various laser bird deterrent devices have emerged in the prior art. For example, CN220966182U discloses a large-scale converter station laser bird deterrent device based on target recognition, which collects data through infrared sensors and visual image sensors, and triggers the laser bird deterrent after target recognition by the control backend. CN116171970A discloses a substation laser bird deterrent device that uses infrared sensors and radar sensors to detect the presence of birds, and controls the laser emission by a microcontroller.

[0005] However, the aforementioned existing technologies have the following shortcomings: 1. Conflict between real-time performance and accuracy: Front-end devices are limited by computing power (usually AI computing power <2TOPS). If a large model is deployed, the inference latency will be >200ms, which cannot meet the real-time deflection requirements of fast-moving targets. If a small model is deployed, the recognition accuracy will be low (<85%), and it will be easy to miss detections.

[0006] 2. Conflict between communication bandwidth and reliability: If all images are uploaded to the cloud, the 4G / 5G signal of the substation may be unstable, and uploading high-definition video streams will consume a lot of data (>1GB per hour), which is uneconomical.

[0007] 3. Poor model adaptability: Substation environments vary greatly with the seasons, lighting, and bird species, making static models unable to adapt, while frequent manual model updates result in high maintenance costs.

[0008] 4. The single method of repelling birds is not adaptable enough: different bird species have different sensitivities to lasers and sounds, so fixed strategies are ineffective, and frequent laser emission will shorten the life of the equipment.

[0009] Therefore, there is an urgent need for an intelligent bird-repelling method that can balance real-time performance, accuracy, communication overhead, and model adaptability. Summary of the Invention

[0010] Based on the technical problems existing in the background technology, this invention proposes a method, system and medium for combined acoustic and optical bird deterrence based on the collaboration of large and small models, which solves the contradiction between real-time performance and accuracy in edge computing scenarios, and improves the bird deterrence effect and equipment lifespan.

[0011] This invention proposes a combined acoustic and optical bird deterrence method based on size model coordination, comprising: Control the pan-tilt-zoom (PTZ) camera to cruise and capture images; The edge computing module runs a front-end small model to perform real-time detection on the image. When the first confidence level of the detection is greater than or equal to the first threshold, it is determined to be a suspected target; the image region of the suspected target is captured and uploaded to the back-end server. The backend server runs a large backend model to verify the received images. If the second confidence level after verification is greater than or equal to the second threshold, the image is confirmed as a bird target. If the second confidence level is greater than the first threshold but less than the second threshold, the final determination is made by combining multi-frame temporal information. Once the bird target is identified, differentiated audio-visual deterrence operations are performed based on the bird species. The backend server performs asynchronous incremental training based on the review results and updates the frontend mini-model through knowledge distillation.

[0012] Furthermore, the front-end small model is a lightweight object detection model, which adopts the MobileNetV3-SSD architecture and is deployed on the NPU unit of the edge computing module for hardware-accelerated inference after being quantized by the HiSilicon NNIE engine INT8.

[0013] Furthermore, the final determination based on the combined temporal multi-frame information specifically involves: extracting M consecutive frames of images; if the second confidence level of 2M / 3 frames is greater than or equal to the first threshold, then the target is determined to be a bird; otherwise, it is determined to be a false alarm.

[0014] Furthermore, the first threshold is adaptively adjusted according to the current zoom level of the gimbal: the first threshold is lowered when the zoom level increases and raised when the zoom level decreases.

[0015] Furthermore, when the magnification factor is greater than 10, the first threshold is reduced to 0.45; when the magnification factor is less than 5, the first threshold is increased to 0.65.

[0016] Furthermore, the differentiated audio-visual deterrence operation based on bird category includes: For small birds, the loudspeaker emits ultrasonic waves; if the target is still present, a single laser beam is fired to scan it. For medium-sized birds, activate the loudspeaker to play falcon calls and two laser beams to set a trajectory for sweeping; For large birds, first control the gimbal to rotate rapidly back and forth and play an alarm sound. If this is ineffective, then launch a multi-beam laser array to sweep.

[0017] Furthermore, the knowledge distillation employs soft-label distillation, with the loss function being... ,in For hard label loss, For soft label loss, These are the weighting coefficients.

[0018] Furthermore, the aforementioned For cross-entropy loss, This represents the KL divergence loss.

[0019] A computer system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0020] A computer-readable storage medium having stored thereon a plurality of computer programs for being invoked by a processor and executing the method as described above.

[0021] The advantages of the combined acoustic and optical bird deterrence method, system, and medium based on large and small model collaboration provided by this invention are as follows: by using the front-end small model accelerated by HiSilicon NPU for rapid screening, the back-end large model for accurate verification, dynamic confidence linkage triggering, and asynchronous incremental distillation updates, the contradiction between real-time performance and accuracy in edge computing scenarios is resolved, thereby improving the bird deterrence effect and equipment lifespan. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the system architecture according to an embodiment of the present invention; Figure 2 A schematic diagram of a laser bird deterrent gimbal camera. Figure 3 For the control method flowchart; Figure 4 This is a schematic diagram of the collaborative recognition process for large and small models. Detailed Implementation

[0023] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] like Figures 1 to 4 As shown, the present invention proposes a combined acoustic and optical bird deterrence method based on size model coordination, comprising: Step 1: Control the PTZ camera to cruise and acquire images; Control the laser bird deterrence PTZ camera to scan according to the preset position sequence and cruise route, stay at each preset position for 3-10 seconds, and rotate the PTZ horizontally at a speed of 15° / s-30° / s.

[0025] Step 2: The edge computing module runs a front-end small model to perform real-time detection on the image. When the detection confidence level is greater than or equal to the first threshold, it is determined to be a suspected target; the image region of the suspected target is captured and uploaded to the back-end server. A high-definition camera captures images at 15-30fps. The edge computing module runs a MobileNetV3-SSD model (input 320×320) quantized with the HiSilicon NNIE engine using INT8 for frame-by-frame detection. The model is trained on the Caffe framework, compiled with INT8 quantization using the HiSilicon RuyiStudio toolchain, and deployed on the NPU unit for hardware-accelerated inference, with a single-frame inference time ≤30ms. A first threshold T1 is set to 0.5-0.6; when the detection confidence level is ≥T1, the target is identified as a suspected bird.

[0026] Non-maximum suppression is applied to suspected targets. The target bounding box with the highest confidence is selected, and the image of the bounding box area is cropped (with a margin of 20% (adjustable)). The image is then uploaded to the backend server via 4G / 5G, along with the azimuth angle, pitch angle, and timestamp of the gimbal.

[0027] Step 3: The backend server runs the backend large model to verify the received image. If the second confidence score after verification is greater than or equal to the second threshold, it is confirmed as a bird target. If the second confidence score is greater than the first threshold and less than the second threshold, the final determination is made by combining the time-series multi-frame information. The backend server calls the YOLOv8x large model (input 640×640) and outputs fine-grained classification results and the second confidence score C2. If C2 ≥ the second threshold T2 (T2=0.85-0.9), it is confirmed as a bird target, and the bird species type label is output.

[0028] If C2 is between T1 and T2 (0.6 ≤ C2 < 0.85), then three consecutive frames are extracted. If C2 ≥ T1 in at least two of these frames, the target is still identified as a bird; otherwise, it is considered a false alarm. Furthermore, T1 is adaptively adjusted based on the current zoom level of the gimbal: T1 decreases to 0.45 when zoom is >10x, and increases to 0.65 when zoom is <5x.

[0029] Step 4: After confirming the bird target, perform differentiated audio-visual deterrence operations based on the bird category; For small birds, control the loudspeaker to emit ultrasonic waves (20-25kHz, 90-100dB) (e.g., 3 seconds), and if the target is still present, emit a single laser beam to scan for a certain period of time (e.g., 1-2 seconds). For medium-sized birds, activate the loudspeaker to play falcon calls (80-95dB) and two laser beams to set a trajectory for sweeping (e.g., a figure-eight sweep, lasting 3-5 seconds). For large birds, first control the gimbal to rotate rapidly back and forth (e.g., 60° / s) and play an alarm sound (100-110dB) (e.g., 5 seconds). If this is ineffective, then launch a multi-beam laser array to scan (e.g., a 55-beam laser array to scan for 5-8 seconds).

[0030] Step 5: The backend server performs asynchronous incremental training based on the review results and updates the frontend mini-model through knowledge distillation.

[0031] Bird images and C2 values ​​confirmed by the large backend model are stored in the backend database. When the cumulative number of newly added positive samples reaches 500-1000, incremental fine-tuning of the large model is triggered. Knowledge distillation (temperature parameter T=4, loss weight α=0.3) is used to compress the knowledge of the large model into a small model. After generating new model parameters, INT8 quantization and compilation are performed again through the HiSilicon RuyiStudio toolchain to generate NNIE wk format model files. The model is pushed to the edge computing module via 4G / 5G every morning, and hot switching is achieved through a dual-caching mechanism (switching time <100ms), realizing silent updates of the frontend model.

[0032] This embodiment uses techniques such as rapid screening of small front-end models accelerated by HiSilicon NPU, precise verification of large back-end models, dynamic confidence linkage triggering, and asynchronous incremental distillation updates to resolve the contradiction between real-time performance and accuracy in edge computing scenarios, thereby improving bird deterrence effects and equipment lifespan.

[0033] Example 1 This embodiment provides a combined acoustic and optical bird deterrence system based on large and small model collaboration, including a laser bird deterrence PTZ camera 100 and a backend server 200.

[0034] like Figure 2 As shown, the laser bird deterrent pan-tilt camera 100 is mounted on a bracket in the substation equipment area, at a height of 3-5 meters. The pan-tilt body 110 adopts an industrial-grade waterproof and dustproof design (IP66), achieving continuous rotation from 0-360° via the horizontal rotation mechanism 111, and tilting from -30° to +90° via the vertical rotation mechanism 112. The laser module 120 uses a green laser (wavelength 532nm, power ≤10mW). The high-definition visible light camera 130 uses a 4-megapixel CMOS sensor and supports 30x optical zoom.

[0035] The edge computing module 140 uses the HiSilicon Hi3403V100 embedded AI SoC chip (unlike the low-computing-power edge chips commonly used in existing technologies (usually <2 TOPS), this embodiment uses the HiSilicon Hi3403V100 high-computing-power chip (10.4 TOPS), providing sufficient hardware resources for the efficient operation of the small front-end model). This chip integrates a quad-core Cortex-A55 processor and the NNIE (Neural Network Inference Engine) neural network inference engine, providing 10.4 TOPS of INT8 computing power and supporting hardware acceleration of deep learning neural networks. The edge computing module 140 deploys the INT8 quantized MobileNetV3-SSD model (small front-end model, 2.3M parameters). The model is trained based on the Caffe framework, and INT8 quantization and compilation are completed through the HiSilicon RuyiStudio toolchain to generate the NNIE wk format model file. Running on the NNIE engine, the single-frame inference time is 25ms. The storage module 160 is a 32GB eMMC, divided into two partitions (partition A stores the currently running model, and partition B stores the model to be updated). The 4G / 5G communication module 150 uses Quectel RM500Q-GL. The loudspeaker module 170 includes a 15W amplifier and a waterproof speaker (maximum SPL 105dB).

[0036] The backend server 200 uses a GPU server (NVIDIA Tesla T4 or NVIDIA® GeForce RTX™ 4090) to deploy a large YOLOv8x model (68M parameters), supporting fine-grained bird classification (50 common substation bird species). The backend server 200 communicates with the laser bird deterrent pan-tilt camera 100 via the communication network 300.

[0037] like Figure 3 and 4 As shown, after system startup, the control system controls the gimbal to cruise at preset positions (6 preset positions, each lasting 5 seconds). The camera captures images at 25fps, and the HiSilicon NPU of the edge computing module 140 runs a small model for detection. When a target with a confidence level ≥ 0.55 is detected, the target image is captured (expanded from 320×320 to 640×640) and uploaded. The backend large model outputs the results within 180ms. If C2 ≥ 0.88, it is confirmed as a bird. If 0.55 ≤ C2 < 0.88, three consecutive frames are buffered. If two of these frames have C2 ≥ 0.55, it is still determined to be a bird. After confirmation, the control system executes a large bird deterrence strategy based on the bird type (e.g., "magpie"): the gimbal rotates rapidly for 5 seconds + an alarm sounds; if ineffective, five laser beams are fired. After the deterrence is completed, the gimbal returns to cruise.

[0038] Example 2: Quantization Deployment Process of HiSilicon NNIE; This example details the quantization deployment process of the front-end small model on HiSilicon NPU.

[0039] S1: Training of a small front-end model; On the PC, a MobileNetV3-SSD model was trained using the Caffe framework and 50,000 labeled images collected at the substation site, outputting a caffemodel file with an accuracy of FP32.

[0040] S2: Model quantization; Using the nnertc tool from the HiSilicon RuyiStudio toolchain, 2000 typical scene images collected at the substation were used as the calibration set to perform INT8 quantization calibration on the FP32 model, calculating the quantization scaling factors and zero points of each layer. HiSilicon NNIE supports quantization of algorithm models in full 8-bit or mixed precision modes. Through a combination of saturated quantization (which requires truncation) and non-saturated quantization strategies, the quantization accuracy loss was controlled to within 3%.

[0041] S3: Model compilation; The quantized front-end small model is compiled using RuyiStudio to generate an NNIE executable wk format model file and corresponding instruction set, which is then deployed to run on the NPU of the edge computing module.

[0042] S4: Inference execution; The edge computing module 140 loads the wk model file and calls the NNIE interface of the HiSilicon MPP media processing platform for hardware-accelerated inference. NNIE provides a dedicated convolutional neural network acceleration engine, supporting the efficient execution of common operators such as convolutional layers, pooling layers, and fully connected layers.

[0043] Example 3: Threshold setting; The edge computing module runs a front-end small model to perform real-time detection on the image. When the first confidence level C1 is greater than or equal to the first threshold T1, it is determined to be a suspected target; the image region of the suspected target is captured and uploaded to the back-end server.

[0044] The first threshold is adaptively adjusted based on the current zoom level of the gimbal: the first threshold decreases when the zoom level increases and increases when the zoom level decreases; specifically: when the zoom level is greater than 10x, the first threshold decreases to 0.45; when the zoom level is less than 5x, the first threshold increases to 0.65.

[0045] Corresponding pseudocode: # Front-end small model detection (HiSilicon NPU acceleration) def frontend_detect(frame, zoom_factor): # Using the HiSilicon NNIE engine for INT8 inference boxes = hisi_nnie_inference(frame, model_file="model.wk", input_size=(320,320)) boxes = nms(boxes, iou_thresh=0.5) # Adaptive threshold T1 = 0.55 if zoom_factor>10: T1 = 0.45 elif zoom_factor<5: T1 = 0.65 if max(boxes.conf) > T1: crop = expand_bbox(frame, boxes[0], margin=0.2) return crop, boxes[0].conf, azimuth, tilt return None The backend server runs a large backend model to verify the received images. If the second confidence level C2 after verification is greater than or equal to the second threshold T2 (T2=0.85-0.9), it is confirmed as a bird target. If the second confidence level is greater than the first threshold and less than the second threshold, the final determination is made by combining the time-series multi-frame information.

[0046] The final determination is made by combining multi-frame temporal information. Specifically, M consecutive frames of images are extracted. If the second confidence level of 2M / 3 frames is greater than or equal to the first threshold, the target is determined to be a bird; otherwise, it is determined to be a false alarm.

[0047] Corresponding pseudocode: # Backend large model verification and dynamic decision-making buffer = [] # Stores C2 for 3 consecutive frames def backend_verify(crop): class_id, conf = yolo_inference(crop, input_size=(640, 640)) T1, T2 = 0.55, 0.88 if conf>= T2: return True, class_id, conf elif T1<= conf <T2: buffer.append((class_id, conf)) if len(buffer) >= 3: high_conf_count = sum(1 for _, c in buffer if c>= T1) if high_conf_count>= 2: # Majority vote determines bird species from collections import Counter majority_class = Counter([c for c, _ in buffer]).most_common(1)[0][0] avg_conf = sum(c for _, c in buffer) / 3 return True, majority_class, avg_conf else: buffer.clear() return False, None, None return False, None, None Example 4: Asynchronous incremental distillation update mechanism.

[0048] The backend server maintains a hard example sample library, which includes the following two categories: Positive samples: images of birds confirmed by the large backend model (C2≥0.88) and images identified as birds by dynamic decision (0.55≤C2<0.88 and passed time-series validation); Negative samples: Images that the large model classifies as non-birds but the small model misclassifies (C2) <T1)。

[0049] When the cumulative number of positive samples increases by 500, incremental fine-tuning of the backend large model is triggered: using the SGD optimizer, learning rate 1e-5 (adjustable), batch size 32 (adjustable), and 100 iterations (adjustable). The fine-tuned backend large model serves as the teacher model, and the current frontend small model serves as the student model, for knowledge distillation. Distillation loss function: ,in For hard-label loss, cross-entropy loss is typically used. For soft label loss, KL divergence loss is typically used. For example, weighting coefficients (Temperature T=4 (adjustable)). The parameters of the distilled small model remain unchanged, but the accuracy is improved by about 5-8 percentage points. The distilled small model is re-quantized and compiled using the HiSilicon RuyiStudio toolchain to generate a new wk model file. After MD5 verification, the backend server pushes it to the second partition of the camera's storage module at 2:00 AM (adjustable) every day and sets an update flag. After the camera detects the flag, it switches the model pointer of the NNIE inference engine from partition A to partition B during the idle time of the next frame. The switching time is 80ms (adjustable) and does not affect continuous detection.

[0050] Example 5: Field experimental data of adaptive threshold and expulsion strategy; The system was deployed at a 220kV substation for 6 months, and the recorded data is shown in Table 1: Table 1

[0051] Among them, the success rate of driving away large birds (magpies and crows) improved most significantly (from 68% to 94%), mainly due to the combination strategy of rapid gimbal rotation and high-decibel alarm. The front-end inference latency decreased from 85ms (CPU inference) to 25ms (HiSilicon NPU hardware acceleration), a reduction of 70%, which significantly improved the real-time response capability of the system.

[0052] Example 6: Computer-readable storage medium This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the methods described in any of the above embodiments. The storage medium includes, but is not limited to, ROM, RAM, hard disk, USB flash drive, SD card, etc.

[0053] This embodiment has the following excellent effects: 1. HiSilicon NPU Hardware Acceleration: Utilizing the INT8 quantization inference capability of the HiSilicon NNIE engine, lightweight models can be run efficiently at the edge (single frame ≤30ms), balancing real-time performance and accuracy.

[0054] 2. Improved recognition accuracy: A two-tier architecture of "rapid screening with a small front-end model + precise verification with a large back-end model" is adopted. For ambiguous regions, multi-frame temporal verification is used. Actual testing shows that the recognition accuracy has increased from 82.3% with a single small model to 97.8%, and the false trigger rate has decreased from 12 times / day to 0.8 times / day. 3. Reduced communication bandwidth consumption: Only cropped suspected target images are uploaded (average 20KB / time), saving 99% of traffic compared to uploading the full image (2MB / time), adapting to the unstable 4G / 5G signal environment in substations.

[0055] 4. Extend equipment lifespan: By reducing invalid laser triggers through large-scale model verification, the average daily number of laser triggers has been reduced from 85 to 22, and the expected lifespan has been extended from 2 years to 5 years.

[0056] 5. Adaptive to environmental changes: The gimbal zoom adaptive threshold adjustment and asynchronous incremental distillation update mechanism enable the model to continuously adapt to changes in season, lighting, and bird species without human intervention.

[0057] 6. Eco-friendly deterrence: Differentiated strategies are adopted for different bird species, with priority given to using sound to deter birds and reduce laser damage, which meets environmental protection requirements.

[0058] Based on the above description of the embodiments, those skilled in the art will understand that the sound-optical combined bird deterrence method, system, and medium based on large-scale model collaboration described in this embodiment can be implemented in pure software or deployed and run on a general-purpose or dedicated computing hardware platform. Based on this essence, the technical solution of this embodiment can be specifically implemented in the form of a software product containing program instructions. This software product can be stored on various non-volatile storage media or directly deployed as a local or cloud service. The program instructions are used to cause computer devices with processing capabilities—including but not limited to personal computers, server clusters, mobile terminals, or other network devices—to execute the steps described in this embodiment.

[0059] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A combined acoustic and optical bird-repelling method based on size-model collaboration, characterized in that, include: Control the pan-tilt-zoom (PTZ) camera to cruise and capture images; The edge computing module runs a front-end small model to perform real-time detection on the image. When the first confidence level of the detection is greater than or equal to the first threshold, it is determined to be a suspected target; the image region of the suspected target is captured and uploaded to the back-end server. The backend server runs a large backend model to verify the received images. If the second confidence level after verification is greater than or equal to the second threshold, the image is confirmed as a bird target. If the second confidence level is greater than the first threshold but less than the second threshold, the final determination is made by combining multi-frame temporal information. Once the bird target is identified, differentiated audio-visual deterrence operations are performed based on the bird species. The backend server performs asynchronous incremental training based on the review results and updates the frontend mini-model through knowledge distillation.

2. The method according to claim 1, characterized in that, The front-end small model is a lightweight object detection model, which adopts the MobileNetV3-SSD architecture and is deployed on the NPU unit of the edge computing module for hardware-accelerated inference after being quantized by the HiSilicon NNIE engine INT8.

3. The method according to claim 1, characterized in that, The final determination based on the combined temporal multi-frame information is as follows: extract M consecutive frames of images, and if the second confidence level of 2M / 3 frames is greater than or equal to the first threshold, it is determined to be a bird target; otherwise, it is determined to be a false alarm.

4. The method according to claim 1, characterized in that, The first threshold is adaptively adjusted according to the current zoom level of the gimbal: the first threshold is lowered when the zoom level increases and raised when the zoom level decreases.

5. The method according to claim 4, characterized in that, When the magnification factor is greater than 10, the first threshold is reduced to 0.45; when the magnification factor is less than 5, the first threshold is increased to 0.

65.

6. The method according to claim 1, characterized in that, The differentiated audio-visual deterrence operation based on bird category includes: For small birds, the loudspeaker emits ultrasonic waves; if the target is still present, a single laser beam is fired to scan it. For medium-sized birds, activate the loudspeaker to play falcon calls and two laser beams to set a trajectory for sweeping; For large birds, first control the gimbal to rotate rapidly back and forth and play an alarm sound. If this is ineffective, then launch a multi-beam laser array to sweep.

7. The method according to claim 1, characterized in that, The knowledge distillation described uses soft-label distillation, with the loss function being: ,in For hard label loss, For soft label loss, These are the weighting coefficients.

8. The method according to claim 7, characterized in that, The For cross-entropy loss, This represents the KL divergence loss.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of computer programs, which are used to be invoked by a processor and to execute the method as described in any one of claims 1-8.

Images