CN-122002134-A - Ship characteristic snapshot method based on track prediction and active zooming

CN122002134ACN 122002134 ACN122002134 ACN 122002134ACN-122002134-A

Abstract

The invention discloses a ship characteristic snapshot method based on track prediction and active zooming, which realizes real-time discovery through a lightweight detection model, combines POS data and an imaging model to calculate geographic coordinates, smoothes tracks by Kalman filtering and predicts future positions, drives a cradle head to adjust Pan/Tilt/Zoom parameters in advance to finish accurate alignment, triggers high-resolution snapshot at the optimal moment, finally recognizes ship names and IMO/MMSI numbers through OCR to form traceable ship characteristic records, and supports automatic comparison and verification with AIS data.

Inventors

LI YONGJUN
Chen Shutu
XU MIAO
LI XIAOFAN
LI MINGHUI

Assignees

江苏欣网视讯软件技术有限公司
南京欣网飞联无人机科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260129

Claims (8)

1. A ship characteristic snapshot method based on track prediction and active zooming is characterized by comprising the following steps: (a) The unmanned aerial vehicle periodically flies according to a preset route, real-time video acquisition is carried out on a monitored water area in a wide-angle mode, and frame-by-frame reasoning is carried out on video frames by adopting a lightweight ship target detection model, so that an initial pixel bounding box and confidence of a ship target are obtained; (b) Based on the initial pixel bounding box, combining longitude, latitude and course angle of the unmanned aerial vehicle platform, internal parameters, direction angle, pitch angle and current focal length of the image acquisition unit, utilizing a small-hole imaging model to reversely solve the real-time longitude and latitude position of the ship under the WGS-84 coordinate system, carrying out track smoothing on the positions of continuous multiframes through a Kalman filter, and estimating real-time navigational speed v and course angle theta of the ship; (c) Calculating the predicted longitude and latitude position P 'of the ship after the delta t time according to the navigational speed v, the heading angle theta and the system delay delta t, and taking the predicted longitude and latitude position P' as an optimal snapshot point; (d) According to the relative geometric relation between the predicted position P' and the mounting position of the camera, the Pan-Tilt control unit calculates a required horizontal corner Pan, a required pitching corner Tilt and a required focal length Zoom, drives a Pan-Tilt motor to perform closed-loop movement to the preset position with the maximum acceleration, and completes active zooming and alignment; (e) When the cradle head is in place, switching to a long focus mode immediately, triggering high-resolution snapshot, and acquiring a ship close-up image containing bow, ship board and ship name character areas; (f) And simultaneously, positioning and identifying a ship identification number through the ship outline and the prior painting position of MMSI numbers to form a ship characteristic record comprising snapshot time, longitude and latitude, the ship name, the identification number and a close-up map, and writing the record into a supervision database.
2. The ship characteristic snapshot method based on track prediction and active zooming of claim 1 is characterized in that a lightweight ship target detection model is YOLOv-s or YOLOv-n, the lightweight ship target detection model is deployed in an edge calculation box, and the reasoning frame rate is more than or equal to 25 FPS.
3. The ship characteristic snapshot method based on track prediction and active zooming as set forth in claim 1, wherein the state vector of the Kalman filter comprises longitude and latitude, speed, heading and acceleration, and the observation noise R and the process noise Q are dynamically adjusted according to the channel level.
4. The ship characteristic snapshot method based on track prediction and active zooming of claim 1 is characterized in that the system delay delta t is obtained through pre-calibration and comprises detection delay 10 ms, cradle head rotation delay less than or equal to 300 ms, zooming delay less than or equal to 200 ms and exposure delay less than or equal to 30 ms.
5. The ship characteristic snapshot method based on track prediction and active zooming of claim 1 is characterized in that the focal length Zoom is automatically determined by a ship pixel proportion threshold beta of a target surface, and the ship name character height is more than or equal to 60 pixels, wherein beta is more than or equal to 15%.
6. The ship characteristic snapshot method based on track prediction and active zooming of claim 1 is characterized in that the OCR character recognition model adopts a CRNN+CTC architecture, training samples comprise ship name character images in inland, coastal and night backlight scenes, and robustness is improved by adopting a data enhancement mode.
7. The ship characteristic snapshot method based on track prediction and active zooming of claim 1 is characterized in that after the step (f), the ship characteristic record is further compared with real-time AIS messages received by an AIS base station, if the identification number is consistent and the ship name similarity is greater than or equal to 95%, the ship characteristic record is marked as verified, and otherwise, manual review is triggered.
8. The ship characteristic snapshot method based on track prediction and active zooming of claim 1 is characterized in that the ship identification number is IMO number or MMSI number.

Description

Ship characteristic snapshot method based on track prediction and active zooming Technical Field The invention belongs to the technical field of dynamic target tracking, and particularly relates to a ship characteristic snapshot method based on track prediction and active zooming. Background Under the background of current marine law enforcement and increasingly stringent water traffic regulations, unmanned-vehicle-based ship snapshot monitoring technology is facing a key bottleneck for transforming from 'coarse-grained monitoring' to 'fine verification'. Along with the improvement of the density of the marine ships and the concealment of illegal operation means, the real-time and accurate acquisition demands on the ship identity information (such as ship name and MMSI number) become more urgent. However, the existing unmanned aerial vehicle monitoring technology still has obvious short plates in the core links of resolution adaptation, target tracking response, environmental robustness, system coordination and the like, and is difficult to meet the high standard requirements of maritime law enforcement on evidence obtaining rigor and traceability. Specific limitations are as follows: 1. basic limitation of recognition accuracy-resolution and zoom-decoupling Current unmanned aerial vehicle monitored control systems generally employ a single focal length wide angle camera for cruise monitoring. While wide-angle lenses can provide a larger field of view, covering a wider range of sea areas, their video resolution is often limited by hardware power consumption and transmission bandwidth. In the wide-angle mode, the proportion of the key identification marks (such as ship name and board number and MMSI) of the ship in the picture is very small, and the key identification marks usually only account for a very low percentage (lower than 0.5%) of the total pixel, so that character details are blurred and the contrast ratio is insufficient. More seriously, existing systems lack active zoom adaptation mechanisms at the time of long-range shooting. The unmanned aerial vehicle cannot adjust the optical focal length in real time according to the motion state and the distance change of the ship, so that characters in the captured image are too small (less than 10 pixels), and the recognition threshold of an OCR (optical character recognition) algorithm is completely exceeded. The mode of blind capture and passive recognition leads to failure of a large amount of key evidence due to unreadable characters, and greatly weakens the integrity of an evidence chain of an illegally operated ship. 2. Response hysteresis, predictive misalignment and control hysteresis of target tracking The motion characteristics of a marine vessel determine its uncertainty in speed and nonlinearity in trajectory. Under the influence of sea waves, wind power and self-maneuvering factors, the speed fluctuation of the ship is large, and the track is difficult to predict. The existing unmanned aerial vehicle monitoring system has obvious hysteresis in target tracking: And the control delay of the cradle head is that the cradle head needs to be continuously fine-tuned to keep the target centered under the condition of sea wave shooting or larger wind speed. However, there are physical limits to the rotational speed and zoom speed of the pan-tilt, and the control algorithm lacks depth fusion of the vessel state, resulting in the vessel easily deviating from the optimal view angle at the instant of snap shots. And the accumulated delay dislocation is that the unmanned aerial vehicle can generate a system accumulated delay which cannot be ignored in the processes of rotating the cradle head, optically zooming and exposing the image. In front of a ship sailing at a high speed, the time delay can cause a remarkable space dislocation of a target at the snapshot moment, namely, a picture is captured but is the historical position of the ship, so that the snapshot image is difficult to correspond to the real ship identity. 3. Adaptability defect of complex environment, robustness is urgently needed to be improved The offshore environment is extremely complex and changeable, and extreme weather and illumination conditions such as back light, night, stormy waves and the like are normal. When the back light shooting is carried out, the ship name and the ship side are often reflected strongly by sunlight, so that the character area is overexposed or completely blocked by light spots. The existing character recognition algorithm is mainly used for training in a uniform illumination environment, has extremely weak feature extraction capability on a high-light-intensity reflection area, and is easy to misjudge. Night shooting is dependent on navigation lights carried by ships or light supplementing lamps of unmanned aerial vehicles. The uneven light source distribution and frequent lamplight flickering can cause the rapid increase of image noise, and the characteri