Bird repelling video target detection adaptive sampling method based on event driving

Abstract

The invention relates to a bird repelling video target detection adaptive sampling method based on event driving, and belongs to the technical field of airport bird repelling detection. The method comprises the following steps: receiving a continuous video stream from video acquisition equipment in real time; setting an event triggering condition based on moving target detection and an event triggering condition based on target recognition; performing real-time analysis on the continuous video stream, capturing a bird repelling event and generating an event signal through an event coding model; the event analysis unit queries a corresponding target sampling rate according to the trigger type and confidence of the event signal; packaging and transmitting a control instruction including the equipment ID, the target sampling rate, and event starting time and predicted ending time determined based on the event signal to the video acquisition equipment; the video acquisition device responds to the control instruction and acquires and generates a key frame sequence; and transmitting the key frame sequence to a target detection unit to execute a target detection task. And the data redundancy and the calculation burden are remarkably reduced while the detection precision is ensured.

Description

Technical Field

[0001] This invention belongs to the field of airport bird detection technology, specifically relating to an event-driven adaptive sampling method for bird detection video targets. Background Technology

[0002] With the expansion of air transport, bird strikes pose increasingly serious challenges to flight safety, airport operations, and ecological environment management. Traditional bird control systems typically employ continuous high-frame-rate video surveillance, indiscriminately collecting and uploading all video data from all scenarios, and then performing bird target identification and trajectory analysis on a backend server. This method has the following drawbacks: Regardless of the presence of bird activity in the scene, the system must run continuously at a fixed high frame rate, resulting in significant consumption of network bandwidth, storage space, and computing resources. Massive amounts of redundant data must be decoded, cached, and inferred by a central server, leading to long processing links and slow processing speeds, making it difficult to meet real-time bird control requirements. The fixed sampling strategy cannot adaptively adjust to changes in lighting, scene complexity, or differences in bird behavior, resulting in missed detections in low-light conditions or during rapid flight maneuvers; simultaneously, it generates numerous false alarms in static or disturbed scenes. When the number of cameras or the addition of high-resolution sensors at the airport increases, the load on the central server increases linearly or even exponentially, making smooth capacity expansion difficult. Existing solutions lack an online feedback mechanism for the relationship between detection results (missed detections, false detections) and sampling parameters, causing the sampling rate to deviate from its optimal value over a long period. Summary of the Invention

[0003] To address the aforementioned problems in the existing technology, this invention provides an event-driven adaptive sampling method for bird deterrence video target detection, comprising: The system receives continuous video streams from a video acquisition device in real time; sets event triggering conditions based on moving target detection and event triggering conditions based on target recognition; performs real-time analysis on the continuous video stream; when any of the event triggering conditions is met, it captures bird-repelling events and generates event signals through an event coding model; and transmits the event signals to the event parsing unit. The event parsing unit queries the corresponding target sampling rate from a preset sampling rate mapping table based on the trigger type and confidence level of the event signal; and encapsulates and transmits the control command containing the device ID, target sampling rate, event start time determined based on the event signal, and expected end time to the video acquisition device. The video acquisition device responds to the control command and, within the time period from the start time of the event to the expected end time, adjusts the video sampling rate to the target sampling rate and acquires and generates a keyframe sequence; transmits the keyframe sequence to the target detection unit; and the target detection unit receives the keyframe sequence and performs a target detection task.

[0004] Specifically, the continuous video stream establishes a streaming media session with the video acquisition device using the RTSP / RTMP protocol, maintaining a TCP / UDP-based transmission link; the received compressed bitstream is hardware-decoded to obtain the original YUV frame sequence, and a circular buffer is enabled at the decoding end to store the most recent frame for inter-frame difference and background modeling.

[0005] Specifically, the event triggering conditions based on moving target detection include: Based on consecutive video frames in a circular buffer, the motion pixel region is calculated using the three-frame difference method; morphological operations and connected component analysis are performed on the motion pixel region, and motion trigger signals are generated according to the ratio of the total area of ​​the effective motion region to the frame area and the displacement of the centroid of the motion region between adjacent frames; the trigger time, the list of bounding boxes of the motion region, and the trigger intensity value are encapsulated as motion event metadata.

[0006] Specifically, the event triggering conditions based on target recognition include: Based on the list of bounding boxes of moving regions generated by the event triggering conditions based on moving target detection, ROI is extracted for each moving region; HOG features are extracted from the ROI and input into a lightweight classifier; based on the bird-like confidence of the ROI and the proportion of the ROI area in the frame area, it is determined whether it contains bird-like motion features and a recognition trigger signal is generated; the trigger time, the bounding box coordinates of the ROI with the highest confidence, and the bird-like feature vector are encapsulated into recognition event metadata.

[0007] Specifically, the event coding model integrates the input motion event metadata and recognition event metadata into continuous events, determines the time continuity based on the time interval and regional overlap rate in the metadata, marks the discrete event metadata as the same event chain based on the continuity determination result, and records the event chain length. Within the same event chain, metadata is merged according to the priority of trigger type characteristics, the event chain length is accumulated, and the latest trigger time is updated; The trigger type features, motion intensity value, confidence features of bird-like confidence, and total bounding box area ratio features are extracted from the integrated single event chain. Event codes for environmental disturbance events, static bird events, and dynamic target events are generated based on feature combinations; a structured signal consisting of the event codes, fusion confidence, aggregated bounding box list, and event time window is output.

[0008] Specifically, the sampling rate mapping table is a two-dimensional decision table, with row indices representing event codes and column indices representing discretized confidence intervals; each cell stores a baseline sampling rate.

[0009] Specifically, the video acquisition device parses the received control command, extracts the device ID, target sampling rate, event start time, and expected end time; performs device ID matching verification, and discards the control command if the command device ID does not match the local ID; Create a dual-buffered acquisition channel: the regular channel maintains the base sampling rate; the event channel is dynamically configured according to the target sampling rate. During the period from the start to the end of the event time window, the video stream is switched to the event channel, and hardware-level frame capture is performed at the target sampling rate; keyframes are extracted to generate a keyframe index table containing timestamp sequences; Specifically, the generation steps of the keyframe sequence within the video acquisition device are as follows: During the event triggering period, the image sensor acquisition parameters are configured according to the target sampling rate in the control command to acquire raw image frame data and synchronously mark time identification information for each frame. Color correction and resolution adaptation processing are performed on the original frame data to make it conform to the input specifications of the target detection unit; A predefined bitrate control model is used to compress the frame sequence, where the upper limit of the bitrate is dynamically calculated based on the target sampling rate. ; Where Bmax is the maximum bitrate, Rbase is the base bitrate, ftarget is the target sampling rate, fbase is the base sampling rate, and Bevent is the additional bitrate increment for the event channel. Generate a metadata index containing a time identifier and frame number for each frame; synchronously write the compressed frame data and metadata index into a temporary storage area, the capacity of which is configured to store complete event window data; The keyframe sequence is fragmented and sent to the target detection unit via a real-time transmission protocol, with the protocol extension header carrying the metadata index; a stream termination signal is sent after the event time window ends; and the actual sampling parameters are fed back to the event parsing unit for system calibration.

[0010] Specifically, the target detection unit performs the following processing steps after receiving the keyframe sequence: The integrity of the keyframe sequence is checked based on the time identifier and frame number in the metadata index. If a frame is missing or out of order, a frame replacement request is sent to the event parsing unit. A pre-trained bird detection model is used to reason frame by frame through the keyframe sequence and output the detection results including category labels, confidence scores, and bounding boxes. Based on the detection results, trajectory association and deduplication are performed to generate continuous bird trajectory segments and calculate trajectory confidence. The trajectory segments, trajectory confidence, and corresponding time windows are encapsulated into bird deterrence alarm messages and pushed to the bird deterrence execution agency through a message queue. The statistical information on missed detections and false detections during the detection process is sent back to the event parsing unit to dynamically update the sampling rate mapping table and the threshold of the lightweight classifier.

[0011] Specifically, the event parsing unit periodically summarizes the actual sampling rate from the video acquisition device, the timestamps of the first and last frames, and the statistics of missed detections / false detections from the target detection unit to form an event-detection closed-loop log; Using the closed-loop log, the baseline sampling rate of each cell in the sampling rate mapping table is updated by an exponential weighted average based on a sliding time window. The update weight is negatively correlated with the event confidence and the false negative rate. When the false negative rate is lower than the preset lower limit and the false positive rate is higher than the preset upper limit within a continuous time window, the baseline sampling rate of the corresponding event code is automatically reduced; conversely, when the false negative rate is higher than the preset upper limit, the baseline sampling rate is increased, and the increase step size is greater than the decrease step size. The updated sampling rate mapping table will be synchronously pushed to all online video capture devices.

[0012] The beneficial effects of this invention are as follows: By employing an event-driven adaptive sampling method, system resource consumption is significantly reduced. The video acquisition device dynamically adjusts the sampling rate only when a bird-scaring event is detected, avoiding the bandwidth and storage pressure caused by continuous high frame rate operation in traditional solutions. Motion triggering and recognition triggering are pushed to the edge, and keyframes are directly pushed to the target detection unit after edge preprocessing, reducing end-to-end latency. A closed-loop feedback mechanism based on event confidence and false negative rate can optimize sampling parameters in real time, ensuring high-precision capture of bird targets in critical scenarios while reducing false alarm rates in static or disturbed scenes. Furthermore, the dual-buffered acquisition channel design ensures parallel processing capabilities for basic monitoring and event response, further improving system stability and reliability. This method not only achieves precise adaptation to the complex environment of airports but also provides scalable technical support for large-scale camera deployment. Attached Figure Description

[0013] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0014] Figure 1 This is a schematic diagram of the structure of an event-driven adaptive sampling method for bird-scaring video target detection according to the present invention. Detailed Implementation

[0015] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0016] Please see Figure 1An event-driven adaptive sampling method for bird-scaring video target detection includes: The system receives continuous video streams from a video acquisition device in real time; sets event triggering conditions based on moving target detection and event triggering conditions based on target recognition; performs real-time analysis on the continuous video stream; when any of the event triggering conditions is met, it captures bird-repelling events and generates event signals through an event coding model; and transmits the event signals to the event parsing unit. The event parsing unit queries the corresponding target sampling rate from a preset sampling rate mapping table based on the trigger type and confidence level of the event signal; and encapsulates and transmits the control command containing the device ID, target sampling rate, event start time determined based on the event signal, and expected end time to the video acquisition device. The video acquisition device responds to the control command and, within the time period from the start time of the event to the expected end time, adjusts the video sampling rate to the target sampling rate and acquires and generates a keyframe sequence; transmits the keyframe sequence to the target detection unit; and the target detection unit receives the keyframe sequence and performs a target detection task.

[0017] Specifically, the continuous video stream establishes a streaming media session with the video acquisition device using the RTSP / RTMP protocol, maintaining a TCP / UDP-based transmission link; the received compressed bitstream is hardware-decoded to obtain the original YUV frame sequence, and a circular buffer is enabled at the decoding end to store the most recent frame for inter-frame difference and background modeling.

[0018] In this embodiment, 4K PTZ cameras are deployed as video acquisition devices at the runway end of the airport. After the system is powered on, each camera establishes a streaming media session with the edge server via RTSP over TCP. The session URL format is: rtsp: / / camera-id:554 / stream1.

[0019] The link layer uses Gigabit Ethernet, and the transport layer simultaneously opens UDP backup port 55443. When the TCP packet loss rate is >5%, it automatically switches to UDP to ensure the continuity of the bitstream.

[0020] The edge server incorporates an NVIDIA Jetson Xavier NX as its hardware decoding unit. After receiving the H.265 compressed bitstream, it uses NVENC / NVDEC hardware decoding to output a 1920×1080@25 fps YUV420 frame sequence. A 1GB circular buffer is allocated at the decoding end, caching the latest 300 frames using a "write pointer-read pointer" mechanism for real-time reading by the three-frame differential and background modeling threads.

[0021] Specifically, the event triggering conditions based on moving target detection include: Based on consecutive video frames in a circular buffer, the motion pixel region is calculated using the three-frame difference method; morphological operations and connected component analysis are performed on the motion pixel region, and motion trigger signals are generated according to the ratio of the total area of ​​the effective motion region to the frame area and the displacement of the centroid of the motion region between adjacent frames; the trigger time, the list of bounding boxes of the motion region, and the trigger intensity value are encapsulated as motion event metadata.

[0022] In this embodiment, the decoding thread sequentially retrieves the latest three grayscale images of F(t-2), F(t-1), and F(t) from the 300-frame circular buffer every 40 ms; Calculate D =|F(t-1)−F(t-2)|; D2=|F(t)−F(t-1)|; take the intersection D=D1∩D2 to obtain the binarized motion pixel map.

[0023] After performing a 3×3 rectangular kernel closure operation (one dilation + one erosion) on the binary graph, connected components are marked using a two-pass scan method; components with an area <90px are removed. 2 The noise zone is defined as the noise zone, and the remaining area is defined as the effective motion zone. Triggering judgment: The area ratio R = Σ effective area area / (1920 × 1080). For the largest effective area, find the centroid C(t), and calculate its displacement from the previous frame's centroid d = ||C(t)−C(t-1)||.

[0024] When R ≥ 0.5% and d ≥ 4px, a motion trigger signal is generated, and the trigger intensity S = R × d.

[0025] The trigger time is the current system timestamp (microsecond-level UTC); the minimum bounding rectangle (x, y, w, h) of the largest valid area is written into the bounding box list; the {trigger time, bounding box list, S} is encapsulated into 32-byte motion event metadata and sent to the event parsing unit via UDP multicast.

[0026] Specifically, the event triggering conditions based on target recognition include: Based on the list of bounding boxes of moving regions generated by the event triggering conditions based on moving target detection, ROI is extracted for each moving region; HOG features are extracted from the ROI and input into a lightweight classifier; based on the bird-like confidence of the ROI and the proportion of the ROI area in the frame area, it is determined whether it contains bird-like motion features and a recognition trigger signal is generated; the trigger time, the bounding box coordinates of the ROI with the highest confidence, and the bird-like feature vector are encapsulated into recognition event metadata.

[0027] In this embodiment, for each motion event metadata generated by the motion detection thread, the N bounding boxes (x, y, w, h) it carries are mapped one by one to the original frame coordinates, and after expanding the edges, they are cropped into ROI sub-images; the sub-images are uniformly scaled to 128×128 pixels, maintaining the aspect ratio and filling with black borders.

[0028] HOG features in the 8×8 cell, 2×2 block, and 9-bin directions are calculated for the scaled ROI to obtain feature vectors, which are then L2 normalized.

[0029] The normalized HOG vector is fed into the SVM model quantized by TensorRT. The model outputs two probabilities: the probability of birds, P. bird and the probability P of non-avian species other .

[0030] Let the area of ​​the current frame be Aframe, and let Pbird ≥ 0.7 if and only if the area of ​​the ROI is A roi ≥0.05%·A frame And P bird / P other When the confidence level is ≥2.5, a recognition trigger signal is generated, with confidence level C=P. bird Encapsulate and identify event metadata. The trigger time is obtained by taking the system's UTC microsecond timestamp; the bounding box (x,y,w,h) of the highest confidence ROI is written into the coordinate field; the multidimensional HOG feature vector is compressed into a 256-dimensional PCA feature and used as the bird-like feature vector; finally, the event metadata is formed and sent to the event coding model via UDP unicast.

[0031] Specifically, the event coding model integrates the input motion event metadata and recognition event metadata into continuous events, determines the time continuity based on the time interval and regional overlap rate in the metadata, marks the discrete event metadata as the same event chain based on the continuity determination result, and records the event chain length. Within the same event chain, metadata is merged according to the priority of trigger type characteristics, the event chain length is accumulated, and the latest trigger time is updated; The trigger type features, motion intensity value, confidence features of bird-like confidence, and total bounding box area ratio features are extracted from the integrated single event chain. Event codes for environmental disturbance events, static bird events, and dynamic target events are generated based on feature combinations; a structured signal consisting of the event codes, fusion confidence, aggregated bounding box list, and event time window is output.

[0032] In this embodiment, the input data consists of 23 pieces of motion event metadata and 7 pieces of recognition event metadata from 4 cameras within the past 3 seconds (75 frames). Example fields: Trigger type: {0 = motion trigger, 1 = recognition trigger}; Trigger time: microseconds UTC; List of bounding boxes: [(x,y,w,h),…]; Confidence level: Exercise intensity value S or bird-like confidence level C; Temporal-spatial continuity determination: Time threshold ΔT = 500ms; Regional overlap threshold θIoU = 0.25; Calculate the time interval dt and maximum IoU for each pair of events: If dt≤ΔT and IoU≥θIoU, then they are marked as the same event chain.

[0033] Ultimately, the 30 discrete events clustered into 3 event chains with lengths of 4, 12, and 14 respectively.

[0034] Intra-chain merging and feature extraction, taking chain 2 (length 12) as an example: Trigger type priority: Recognition trigger > Motion trigger → Set to 1; Fusion confidence F = max(C); Motion intensity M = max(S); The total bounding box area R = combined box area / frame area; Taking the quadruple [T,F,M,R] mapped via a lightweight gradient boosting tree as an example, the structured output is as follows: The output code is 01 (static bird event), with a probability of p=0.91; Event time window: first trigger 14:32:15.200, last trigger 14:32:17.080; Aggregated bounding box: Joint minimum bounding rectangle (1230, 480, 180, 120).

[0035] The structured signal output JSON format is as follows: { "camera_id": "CAM-03", "event_code": 01, "confidence": 0.91, "bbox": [1230, 480, 180, 120], "time_window": [1688101935200000, 1688101937080000] } The event is pushed to the event parsing unit via MQTT QoS1.

[0036] Specifically, the sampling rate mapping table is a two-dimensional decision table, with row indices representing event codes and column indices representing discretized confidence intervals; each cell stores a baseline sampling rate.

[0037] Specifically, the video acquisition device parses the received control command, extracts the device ID, target sampling rate, event start time, and expected end time; performs device ID matching verification, and discards the control command if the command device ID does not match the local ID; Create a dual-buffered acquisition channel: the regular channel maintains the base sampling rate; the event channel is dynamically configured according to the target sampling rate. During the period from the start to the end of the event time window, the video stream is switched to the event channel, and hardware-level frame capture is performed at the target sampling rate; keyframes are extracted to generate a keyframe index table containing timestamp sequences; Specifically, the generation steps of the keyframe sequence within the video acquisition device are as follows: During the event triggering period, the image sensor acquisition parameters are configured according to the target sampling rate in the control command to acquire raw image frame data and synchronously mark time identification information for each frame. Color correction and resolution adaptation processing are performed on the original frame data to make it conform to the input specifications of the target detection unit; A predefined bitrate control model is used to compress the frame sequence, where the upper limit of the bitrate is dynamically calculated based on the target sampling rate. ; Where Bmax is the maximum bitrate, Rbase is the base bitrate, ftarget is the target sampling rate, fbase is the base sampling rate, and Bevent is the additional bitrate increment for the event channel. Generate a metadata index containing a time identifier and frame number for each frame; synchronously write the compressed frame data and metadata index into a temporary storage area, the capacity of which is configured to store complete event window data; The keyframe sequence is fragmented and sent to the target detection unit via a real-time transmission protocol, with the protocol extension header carrying the metadata index; a stream termination signal is sent after the event time window ends; and the actual sampling parameters are fed back to the event parsing unit for system calibration.

[0038] Specifically, the target detection unit performs the following processing steps after receiving the keyframe sequence: The integrity of the keyframe sequence is checked based on the time identifier and frame number in the metadata index. If a frame is missing or out of order, a frame replacement request is sent to the event parsing unit. A pre-trained bird detection model is used to reason frame by frame through the keyframe sequence and output the detection results including category labels, confidence scores, and bounding boxes. Based on the detection results, trajectory association and deduplication are performed to generate continuous bird trajectory segments and calculate trajectory confidence. The trajectory segments, trajectory confidence, and corresponding time windows are encapsulated into bird deterrence alarm messages and pushed to the bird deterrence execution agency through a message queue. The statistical information on missed detections and false detections during the detection process is sent back to the event parsing unit to dynamically update the sampling rate mapping table and the threshold of the lightweight classifier.

[0039] Specifically, the event parsing unit periodically summarizes the actual sampling rate from the video acquisition device, the timestamps of the first and last frames, and the statistics of missed detections / false detections from the target detection unit to form an event-detection closed-loop log; Using the closed-loop log, the baseline sampling rate of each cell in the sampling rate mapping table is updated by an exponential weighted average based on a sliding time window. The update weight is negatively correlated with the event confidence and the false negative rate. When the false negative rate is lower than the preset lower limit and the false positive rate is higher than the preset upper limit within a continuous time window, the baseline sampling rate of the corresponding event code is automatically reduced; conversely, when the false negative rate is higher than the preset upper limit, the baseline sampling rate is increased, and the increase step size is greater than the decrease step size. The updated sampling rate mapping table will be synchronously pushed to all online video capture devices.

[0040] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An event-driven adaptive sampling method for bird-scaring video target detection, characterized in that, include: Receives continuous video streams from video capture devices in real time; Set event triggering conditions based on moving target detection and event triggering conditions based on target recognition; The continuous video stream is analyzed in real time. When any of the event triggering conditions are met, bird deterrence events are captured and event signals are generated through an event coding model. The event signal is transmitted to the event parsing unit; The event parsing unit queries the corresponding target sampling rate from a preset sampling rate mapping table based on the trigger type and confidence level of the event signal; and encapsulates and transmits the control command containing the device ID, target sampling rate, event start time determined based on the event signal, and expected end time to the video acquisition device. The video acquisition device responds to the control command and, within the time period from the start time of the event to the expected end time, adjusts the video sampling rate to the target sampling rate and acquires and generates a keyframe sequence; transmits the keyframe sequence to the target detection unit; and the target detection unit receives the keyframe sequence and performs a target detection task.

2. The method according to claim 1, characterized in that, The continuous video stream establishes a streaming media session with the video acquisition device using the RTSP / RTMP protocol, maintaining a TCP / UDP-based transmission link; the received compressed bitstream is hardware-decoded to obtain the original YUV frame sequence, and a circular buffer is enabled at the decoding end to store the most recent frame for inter-frame difference and background modeling.

3. The method according to claim 1, characterized in that, The event triggering conditions based on moving target detection specifically include: Based on consecutive video frames in a circular buffer, the motion pixel region is calculated using the three-frame difference method; morphological operations and connected component analysis are performed on the motion pixel region, and motion trigger signals are generated according to the ratio of the total area of ​​the effective motion region to the frame area and the displacement of the centroid of the motion region between adjacent frames; the trigger time, the list of bounding boxes of the motion region, and the trigger intensity value are encapsulated as motion event metadata.

4. The method according to claim 1, characterized in that, The event triggering conditions based on target recognition specifically include: Based on the list of bounding boxes of moving regions generated by the event triggering conditions based on moving target detection, ROI is extracted for each moving region; HOG features are extracted from the ROI and input into a lightweight classifier; based on the bird-like confidence of the ROI and the proportion of the ROI area in the frame area, it is determined whether it contains bird-like motion features and a recognition trigger signal is generated; the trigger time, the bounding box coordinates of the ROI with the highest confidence, and the bird-like feature vector are encapsulated into recognition event metadata.

5. The method according to claim 1, characterized in that, The event coding model integrates the input motion event metadata and recognition event metadata into continuous events, determines the temporal continuity based on the time interval and regional overlap rate in the metadata, marks the discrete event metadata as the same event chain based on the continuity determination result, and records the event chain length. Within the same event chain, metadata is merged according to the priority of trigger type characteristics, the event chain length is accumulated, and the latest trigger time is updated; The trigger type features, motion intensity value, confidence features of bird-like confidence, and total bounding box area ratio features are extracted from the integrated single event chain. Event codes for environmental disturbance events, static bird events, and dynamic target events are generated based on feature combinations; a structured signal consisting of the event codes, fusion confidence, aggregated bounding box list, and event time window is output.

6. The method according to claim 1, characterized in that, The sampling rate mapping table is a two-dimensional decision table, with row indexes representing event codes and column indexes representing discretized confidence intervals; each cell stores a baseline sampling rate.

7. The method according to claim 1, characterized in that, The video acquisition device parses the received control command, extracts the device ID, target sampling rate, event start time, and expected end time; performs device ID matching verification, and discards the control command if the command device ID does not match the local ID. Create a dual-buffered acquisition channel: the regular channel maintains the base sampling rate; the event channel is dynamically configured according to the target sampling rate. During the period from the start to the end of the event time window, the video stream is switched to the event channel, and hardware-level frame capture is performed at the target sampling rate; Extract keyframes to generate a keyframe index table containing timestamp sequences.

8. The method according to claim 1, characterized in that, The keyframe sequence is generated within the video acquisition device in the following steps: During the event triggering period, the image sensor acquisition parameters are configured according to the target sampling rate in the control command to acquire raw image frame data and synchronously mark time identification information for each frame. Color correction and resolution adaptation processing are performed on the original frame data to make it conform to the input specifications of the target detection unit; A predefined bitrate control model is used to compress the frame sequence, where the upper limit of the bitrate is dynamically calculated based on the target sampling rate. ； Among them, B max For the maximum bitrate, R base Based on the base bitrate, f target f is the target sampling rate. base Based on the sampling rate, B event Add bitrate increments to the event channel; Generate a metadata index containing a time identifier and frame number for each frame; synchronously write the compressed frame data and metadata index into a temporary storage area, the capacity of which is configured to store complete event window data; The keyframe sequence is fragmented and sent to the target detection unit via a real-time transmission protocol, with the protocol extension header carrying the metadata index; a stream termination signal is sent after the event time window ends; and the actual sampling parameters are fed back to the event parsing unit for system calibration.

9. The method according to claim 1, characterized in that, The target detection unit performs the following processing steps after receiving the keyframe sequence: The integrity of the keyframe sequence is checked based on the time identifier and frame number in the metadata index. If a frame is missing or out of order, a frame replacement request is sent to the event parsing unit. A pre-trained bird detection model is used to reason frame by frame through the keyframe sequence and output the detection results including category labels, confidence scores, and bounding boxes. Based on the detection results, trajectory association and deduplication are performed to generate continuous bird trajectory segments and calculate trajectory confidence. The trajectory segments, trajectory confidence, and corresponding time windows are encapsulated into bird deterrence alarm messages and pushed to the bird deterrence execution agency through a message queue. The statistical information on missed detections and false detections during the detection process is sent back to the event parsing unit to dynamically update the sampling rate mapping table and the threshold of the lightweight classifier.

10. The method according to claim 1, characterized in that, The event parsing unit periodically summarizes the actual sampling rate from the video acquisition device, the timestamps of the first and last frames, and the statistics of missed detections / false detections from the target detection unit to form an event-detection closed-loop log. Using the closed-loop log, the baseline sampling rate of each cell in the sampling rate mapping table is updated by an exponential weighted average based on a sliding time window. The update weight is negatively correlated with the event confidence and the false negative rate. When the false negative rate is lower than the preset lower limit and the false positive rate is higher than the preset upper limit within a continuous time window, the baseline sampling rate of the corresponding event code is automatically reduced. Conversely, when the false negative rate is higher than the preset upper limit, the baseline sampling rate is increased, and the increase step size is greater than the decrease step size. The updated sampling rate mapping table will be synchronously pushed to all online video capture devices.

Images