Method and device for cooperatively tracking target through long-focus lens and short-focus lens

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a method and device for cooperatively tracking a target through a long-focus lens and a short-focus lens. The device comprises a dual-optical-path system unit, a synchronous imaging unit, a tracking decision unit and a tracking execution unit. The tracking decision-making unit decides selection of tracking video data and transition and switching of a tracking algorithm according to the target size information; the tracking execution unit is used for executing a tracking algorithm, and the corresponding tracking execution algorithm is adopted for sensor data of the long-focus light path and the short-focus light path. An end-to-end architecture is adopted in the deep learning algorithm for double-path overlapping tracking, cross-scale alignment of feature layers is achieved, and switching decision errors are avoided. And through accurate size feature analysis, tracking state switching is driven, transition smoothness is ensured by adopting overlapping tracking and confidence fusion, and a complex scene is adapted by combining dynamic parameter adjustment.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image processing, and particularly relates to a method and device for long-focus and short-focus lens cooperative tracking of a target. BACKGROUND

[0002] In a target detection application scenario, functions such as target detection, identification and tracking need to be implemented, and a continuous zoom lens or a multi-view lens is usually selected. The continuous zoom lens realizes stepless adjustment of focal length by moving a plurality of lens groups along the optical axis, such as a wide range of 10mm-500mm or 50mm-1000mm, and the structure includes a variable magnification group, a compensation group and a fixed group, which needs a precise mechanical structure to support a complex motion trajectory. The multi-view lens provides a plurality of fixed focal lengths, such as 20mm / 60mm and 24° / 14°, and realizes field of view jumping by switching lens group positions or splitting light, and the structure is simplified, and double focus switching can be realized by moving a single lens.

[0003] Typical advantages of the continuous zoom lens include: (1) strong flexibility: continuous adjustment of focal length, suitable for complex dynamic scenes; (2) high imaging consistency: seamless connection of images during zooming, avoiding target loss; (3) long-distance detail resolution: supporting ultra-long focal length (such as 500mm), with a detection distance of several kilometers. Typical disadvantages: (1) complex system: precise mechanical structure, vulnerable to vibration affecting stability; (2) high cost: the selling price can be 2-5 times that of a dual-view lens, and maintenance is difficult; (3) switching delay: motor-driven zooming takes hundreds of milliseconds, which is weaker than the millisecond-level response of a dual-view lens.

[0004] Typical advantages of the multi-view lens include: (1) fast switching: physical / electronic switching can be completed within 1 second; (2) compact structure: simplified lens group, such as 5 lenses for double focus, small size and anti-vibration; (3) cost advantage: mass production cost is more than 30% lower, suitable for budget-sensitive scenarios. Disadvantages: (1) limited flexibility: only two or more focal lengths, unable to adapt to continuous observation requirements; (3) field of view transition jumping: target may be lost during switching, which needs to be compensated by an algorithm; (3) weak long-focus capability: usually focal length ≤100mm, weak in long-distance detail capture compared with the zoom lens.

[0005] The application scenarios of photoelectric imaging systems of aircraft loads such as missiles or unmanned aerial vehicles are dynamic wide-area tracking and super-long-distance observation, which need to lock the moving target in real time and realize continuous tracking of the target. In order to increase the distance of the detected target, a long-focus lens with a small field of view is usually selected for the optical system. When the distance is quickly close to the target, the target is easily lost from the field of view or only part of the target can be seen due to the limitation of the field of view size, which causes the target tracking to fail, the precision deviation of the strike target is too large, and the enemy target cannot be caused fatal damage, which has been the pain point of the terminal photoelectric guidance. In addition, it is difficult to realize the requirements of low cost, anti-vibration and zero delay whether a continuous zoom lens or a multi-field lens is used. SUMMARY

[0006] In view of the technical deficiencies of the existing target detection and tracking technology, the application provides a method and device for long-focus and short-focus lens cooperative tracking of a target, which includes a complete double optical path real-time fusion tracking scheme. Through the combination of hardware optical path alignment and software adaptive tracking, continuous stable tracking in the whole distance is realized.

[0007] The following is a detailed technical solution.

[0008] The application first provides a device for long-focus and short-focus lens cooperative tracking of a target, which includes a double optical path system unit, a synchronous imaging unit, a tracking decision unit and a tracking execution unit.

[0009] The double optical path system unit adopts a beam splitter prism to divide the incident light into two paths, one long-focus path and the other short-focus path, or adopts a separate design of long-focus and short-focus lenses, reduces the error in the center direction of the field of view by parallelizing the optical axes of the lenses, and ensures that the center field of view deviation of the double paths is less than 3 pixels at 4K resolution through an optical focus alignment controller.

[0010] The synchronous imaging unit includes a double sensor and an image processor. The double sensor adopts a global shutter visible light CMOS, or adopts an infrared SENSOR, or adopts sensors of different wavebands. The RAW data of the double sensors is triggered and processed synchronously through a single FPGA hardware chip. The image processor adopts the RAW data of the double sensors for ISP processing.

[0011] The tracking decision unit decides the selection of tracking video data and the transition and switching of tracking algorithms according to target size information.

[0012] The tracking execution unit is used to execute the tracking algorithm. The sensor data of the long-focus and short-focus optical paths adopts the corresponding tracking execution algorithm.

[0013] The application also provides a method for long-focus and short-focus lens cooperative tracking of a target using the above device, which specifically includes the following aspects.

[0014] I. Cooperative tracking algorithm

[0015] Long focal path selection KCF correlation filter tracker, short focal path selection SiamRPN tracker;

[0016] Firstly, the long focal path extracts phase features, anti-motion blur

[0017] ;

[0018] Then, the short focal path extracts texture and color features

[0019] ;

[0020] Finally, cross-scale feature fusion

[0021] where α = 0.7 ~ 0.9 is an empirical value;

[0022] II. Switching control based on target size

[0023] The size feature description adopts three key indicators:

[0024] The pixel coverage is the target pixel area divided by the pixel area of the image frame:

[0025]

[0026] The boundary distance ratio is the minimum value of the ratio of the target distance to the left boundary of the image to the distance to the right boundary of the image and the ratio of the target distance to the upper boundary of the image to the distance to the lower boundary of the image:

[0027]

[0028] The aspect ratio anomaly degree is the absolute value of the difference between the target pixel aspect ratio and the true aspect ratio of the target:

[0029]

[0030] where , are the true physical length and width;

[0031] When C < 5% or B > 0.3, the target size is a very small target, the lens uses the long focal path, and the phase correlation tracking strategy is adopted;

[0032] When 5% ≤ C ≤ 30% and B ≥ 0.15, the target size is a medium target, the lens uses the long focal path, and the feature fusion tracking strategy is adopted;

[0033] When 25% ≤ C ≤ 35%, the target size is a critical target, the lens uses both the long focal path and the short focal path, and the overlapping tracking strategy is adopted;

[0034] When C > 30% or B < 0.15, the target size is large target, the lens uses a short focal path, and a local feature tracking strategy is adopted;

[0035] When R > 1.5 and C > 20%, the target size is a distorted target, the lens uses a short focal path, and a forced switching tracking omission is adopted;

[0036] III. Decision-making process of long-short focal lens switching

[0037] 1) The aircraft is getting closer and closer to the target, switching from long focal length to short focal length, which is divided into four stages:

[0038] Stage one: switching trigger

[0039] The process of switching trigger is as follows:

[0040] The process of switching trigger is as follows:

[0041] (1) Long focal length tracker enables state trigger decision layer; for example, report C = 28% for 5 consecutive frames, that is, C > 25% for 5 consecutive frames.

[0042] (2) The decision layer triggers the short focal length system to start pre-tracking;

[0043] (3) The short focal length system triggers the coordinate converter to request target coordinates H_matrix;

[0044] (4) The coordinate converter converts the target coordinates and returns the homography matrix to adapt to the short focal length system;

[0045] (5) The short focal length system triggers the target positioning function, that is, enters the overlapping tracking phase of long focal length tracking and short focal length tracking.

[0046] Stage two: overlapping tracking period (300ms critical period)

[0047] When the video frame rate is 60fps, the overlapping tracker is 300ms, and the following tracking is performed for each frame of image.

[0048] (1) Double parallel tracking, calculating the tracking box bbox of long focal length tracking and short focal length initialization

[0049] Long focal length tracking

[0050] Short focal length initialization

[0051] (2) Confidence fusion

[0052] Long focal length confidence

[0053] Short focal length confidence .

[0054] (3) Weighted output

[0055] Value parameter , value 0→1 gradual change;

[0056] Long-focus weight

[0057] Short-focus weight

[0058]

[0059] (4) Abnormality detection

[0060] Detect whether the tracking box feature is abnormal , if abnormal, recalibrate the target .

[0061] Stage 3: Switching completion determination

[0062] The switching is completed when any of the following conditions is met:

[0063] (1) Short-focus confidence is greater than 0.85 for 5 consecutive frames;

[0064] (2) Weighting factor alpha is greater than 0.7;

[0065] (3) Long-focus feature degradation (gradient variance decreases by 30%).

[0066] Stage 4: Long-focus resource release

[0067] (1) Keep the long-focus detector, run at 1Hz low frequency;

[0068] (2) Store the target feature template;

[0069] (3) Release the tracker resource.

[0070] 2), size change special case processing

[0071] Emergency switching triggered by sudden approach of the target, with two types of conditions, one is long-focus boundary distance ratio ; the second is the aspect ratio abnormality . The processing measure is to start short-focus tracking immediately, and directly switch without skipping the overlap period.

[0072] 3), adaptive optimization of control parameters

[0073] (1) Dynamic threshold adjustment

[0074]

[0075] Among them, It is the target's motion speed (pixels / frame). It is the target acceleration. , These are all empirical coefficients.

[0076] (2) Scenario-based hysteresis control

[0077] No. Scene Type Long→Short Threshold Short→Long Threshold Hysteresis Interval 1 Uniform Motion 28% 5% 5% 2 High-Speed Maneuver 25% 7% 8% 3 Frequent Occlusion 32% 4% 3% 4 Low Contrast 26% 6% 6%

[0078] IV. Ensuring Continuity of Tracking

[0079] The measures to ensure continuous performance tracking are divided into pre-handover diagnostics and post-handover verification, as detailed below:

[0080] (1) Pre-switching diagnosis

[0081] Before switching, pre-switching diagnostic measures are added. First, it is determined whether the target size has reached the threshold. Second, the complexity of the target background is analyzed. If the target size and background are simple, the switching is started immediately. Otherwise, the stability of the target features is further determined. If the features are stable, the switching is performed normally. If the features are unstable, the switching is delayed and the target features are strengthened.

[0082] (2) Verification after switching

[0083] Geometric consistency check:

[0084] Feature similarity verification:

[0085] Motion continuity detection: .

[0086] The device for collaborative target tracking using telephoto and short-focal-length lenses of this invention eliminates parallax through physical optical axis alignment, resolves scale jump issues through software adaptive switching, and ensures tracking continuity with a trajectory prediction algorithm. The key to collaborative target tracking using telephoto and short-focal-length lenses lies in addressing field-of-view differences and scale jumps to achieve seamless relay tracking. This invention employs an end-to-end architecture for dual-path overlapping tracking deep learning, achieving cross-scale alignment of feature layers and avoiding switching decisions. Precise size feature analysis drives state switching, overlapping tracking and confidence fusion ensure smooth transitions, and dynamic parameter adjustments adapt to complex scenarios. Attached Figure Description

[0087] Fig. 1 This is a schematic diagram of the device logic.

[0088] Fig. 2 This is a flowchart of the adaptive tracking switching process between telephoto and short telephoto lenses.

[0089] Fig. 3 Select a tracker. Detailed Implementation

[0090] The long-focus and short-focus lens device of the application cooperates to track the target, and the core components include three parts:

[0091] 1. Dual optical path system unit

[0092] (1) Common-aperture light splitting system (optional scheme one)

[0093] The incident light is split into two paths by a light splitting prism: a long-focus path, for example, 80% light flux, 1000mm focal length, FOV=2°. A short-focus path, for example, 20% light flux, 35mm focal length, FOV=60°; Through optical axis alignment control, the optical axis alignment technology is adopted, and the short-focus light path is adjusted through a micro-displacement piezoelectric ceramic, to ensure that the center field deviation of the dual paths is less than 3 pixels @ 4K resolution.

[0094] (2) Parallel dual optical path system (optional scheme two)

[0095] The long-focus and short-focus lens are designed in a separate manner, and the optical axis parallelism is adopted to reduce the center direction error of the field of view; the optical axis alignment tool is adopted to adjust the assembly process of the structure, to ensure that the center field deviation of the dual paths is less than 3 pixels @ 4K resolution.

[0096] 2. Synchronous imaging unit

[0097] (1) Synchronous sensor

[0098] There are multiple schemes for the dual-path sensor: scheme one: both use a global shutter visible light CMOS, for example, 1920×1080 @ 60fps; time synchronization error ≤100μs; scheme two: both use an infrared sensor, for example, 1280×1024 @ 30fps; time synchronization error ≤200μs; scheme three: use sensors of different wavebands, for example, use a global shutter visible light CMOS, 2560×1440 @ 50fps; use a non-cooled infrared detector, 640×512 @ 50fps; time synchronization error ≤5ms.

[0099] The RAW data of the dual paths is triggered and processed by a single FPGA hardware chip.

[0100] (2) Synchronous image processing

[0101] The RAW data of the dual-path sensor is input into a field programmable logic device FPGA, for ISP processing.

[0102] 3. Tracking decision and tracking execution unit

[0103] (1) Tracking decision unit

[0104] The tracking decision unit decides the selection of the tracking video data and the transition and switching of the tracking algorithm according to the target size information.

[0105] (2) The tracking execution unit

[0106] The tracking execution unit is a module for executing the tracking algorithm, and the sensor data of the long-focus and short-focus optical paths adopts corresponding tracking execution algorithms.

[0107] The method for tracking a target by a long-focus and short-focus lens, the long-focus and short-focus dual-lens cooperative tracking algorithm and the size-driven switching control are the core contents of the present application, and are described as follows:

[0108] 1 Algorithm overall architecture

[0109] A three-layer closed-loop control system is adopted: (1) a target size analysis layer; (2) a lens switching decision layer; and (3) a tracking continuity guarantee layer.

[0110] 1.1 Cooperative tracking algorithm

[0111] 1.1.1 Tracker design

[0112] The long-focus selects a KCF correlation filter tracker, which is suitable for small targets;

[0113] The short-focus selects a SiamRPN tracker, which is suitable for large target deformation.

[0114] 1.1.2 Hybrid feature extraction

[0115] First, the long-focus path extracts phase features, which are resistant to motion blur

[0116]

[0117] Then, the short-focus path extracts texture and color features

[0118]

[0119] Finally, cross-scale feature fusion

[0120]

[0121] Wherein, alpha=0.7~0.9 is an empirical value.

[0122] 1.2 Target size analysis layer

[0123] Switching control mechanism based on target size

[0124] 1.2.1 Size feature definition

[0125] Three key indicators of the size feature description:

[0126] (1) Pixel coverage is the area of the target pixel divided by the area of the image frame pixel:

[0127]

[0128] (2) Border distance ratio is the ratio of the distance from the target to the left border of the image and the distance to the right border of the image, and the ratio of the distance from the target to the top border of the image and the distance to the bottom border of the image, and the minimum value of the two ratios:

[0129]

[0130] (3) Aspect ratio anomaly is the absolute value of the difference between the aspect ratio of the target pixel and the true aspect ratio of the target

[0131]

[0132] wherein , are the true physical length and width.

[0133] 1.2.2 Size-lens mapping relationship

[0134] No. Target Size State Feature Range Recommended Lens Tracking Strategy 1 Distant C < 5% or B > 0.3 Long Phase Correlation Tracking 2 Optimal 5% ≤ C ≤ 30% and B ≥ 0.15 Long Feature Fusion Tracking 3 Transition 25%≤C≤35% Dual Lens Overlap Tracking 4 Close C > 30% or B < 0.15 Short Local Feature Tracking 5 Distorted R > 1.5 and C > 20% Short Forced Switch

[0135] 1.3 Lens switching decision layer

[0136] Decision process of long-short lens switching

[0137] 1.3.1 Switching stage description

[0138] Taking long focus to short focus as an example, it is divided into four stages and described as follows:

[0139] Stage 1: Switching trigger

[0140] The process of switching trigger is as follows:

[0141] (1) The long-focus tracker enables the decision layer; for example, reports that C=28% for 5 consecutive frames, i.e. C>25% for 5 consecutive frames.

[0142] (2) The decision layer triggers the short-focus system to start pre-tracking;

[0143] (3) The short-focus system triggers the coordinate converter to request target coordinates H_matrix;

[0144] (4) The coordinate converter converts the target coordinates and returns the homography matrix, which is adapted to the short-focus system;

[0145] (5) The short-focus system triggers the target positioning function, i.e. enters the overlapping tracking phase of long-focus tracking and short-focus tracking.

[0146] Stage 2: Overlapping tracking period (300 ms critical period)

[0147] Video frame rate is 60fps, overlap tracker 300ms, and then track each frame of image.

[0148] (1) Double parallel tracking, calculate long tracking and short initialization tracking box bbox

[0149] Long tracking

[0150] Short initialization

[0151] (2) Confidence fusion

[0152] Long confidence

[0153] Short confidence .

[0154] (3) Weighted output

[0155] Value parameters , Value 0→1 gradual change;

[0156] Long weight

[0157] Short weight

[0158]

[0159] (4) Abnormal detection

[0160] Detect if the tracking box feature is abnormal , If abnormal, re-calibrate target .

[0161] Stage 3: Switching completion determination

[0162] Any of the following conditions is met to complete the switching:

[0163] (1) Short confidence is greater than 0.85 for 5 consecutive frames;

[0164] (2) Weighting factor alpha is greater than 0.7;

[0165] (3) Long appears feature degradation (gradient variance drops 30%).

[0166] Stage 4: Long resource release

[0167] (1) Keep long detector, 1Hz low frequency operation;

[0168] (2) Store target feature template;

[0169] (3) Release tracker resources.

[0170] 1.3.2 Size change special case processing

[0171] The target suddenly approaches the emergency switching, and the triggering condition has two types, one is the long-focus boundary distance ratio ; The second is the aspect ratio anomaly degree . The processing measure is to immediately start short-focus tracking and directly switch by skipping the overlap period.

[0172] 1.3.3 Adaptive optimization of control parameters

[0173] (1) Dynamic threshold adjustment

[0174]

[0175] Wherein, is the target motion speed (pixels / frame), is the target acceleration, , are empirical coefficients.

[0176] (2) Scene-based hysteresis control

[0177] No. Scene Type Long→Short Threshold Short→Long Threshold Hysteresis Interval 1 Uniform Motion 28% 5% 5% 2 High-Speed Maneuver 25% 7% 8% 3 Frequent Occlusion 32% 4% 3% 4 Low Contrast 26% 6% 6%

[0178] 1.4 Tracking continuity guarantee layer

[0179] The tracking continuity performance guarantee measures are divided into pre-switching diagnosis and post-switching verification, which are specifically described as follows:

[0180] (3) Pre-switching diagnosis

[0181] The pre-switching diagnosis measure is added before switching, first, it is judged whether the target size reaches the threshold value; Secondly, the target background complexity is analyzed, if the target size background is simple, the switching is immediately started; Otherwise, the target feature stability is judged, if the feature is stable, the normal switching is performed; If it is unstable, the switching is delayed, and the target feature is strengthened.

[0182] (2) Post-switching verification

[0183] Geometric consistency check:

[0184] Feature similarity verification:

[0185] Motion continuity detection:

[0186] The present application also has the following key points:

[0187] 1. Light focus space alignment technology

[0188] (1) Dynamic calibration compensation

[0189] A chessboard calibration plate is arranged at the optical focal plane, and mapping is achieved by the following matrix:

[0190]

[0191] Wherein, K is the camera intrinsic parameter, and R is the rotation matrix.

[0192] (2) Real-time distortion correction

[0193] A bicubic spline interpolation algorithm is used to establish a deformation field grid based on the long-focus image, and the calculation complexity is controlled at 15ms / frame.

[0194] 2. Adaptive lens switching strategy

[0195] (1) Scale determination criterion:

[0196] First, define the target tracking frame lens_switch(target_bbox) as the basis for lens selection; then, calculate the target pixel ratio

[0197]

[0198] If , the target occupies the field of view > 30%, switch to the short-focus lens;

[0199] If , the target occupies the field of view < 5%, switch to the long-focus lens;

[0200] Otherwise, keep the current state.

[0201] 3. Cross-scale target tracking algorithm

[0202] (1) Feature fusion tracker, combining long-focus path features and short-focus path features for feature fusion tracking;

[0203] (2) Scale space prediction uses the DSST (Discriminative Scale Space Tracker) model to achieve scale change rate prediction error < 8%.

[0204] 4. Single-chip dual-optical path image and algorithm processing

[0205] (1) FPGA single-chip synchronous clock processing long-focus optical path data and short-focus optical path data ISP algorithm, realizing zero-delay dual-path imaging;

[0206] (2) Soc single-chip implementation of synchronous processing long focal length optical path data and short focal length optical path data tracking algorithm; Realize target size analysis, lens switching decision and tracking continuity guarantee; Low cost, real-time, smooth and stable target tracking switching; Realize hybrid feature extraction, same characteristics of Forbidden City feature fusion tracker adapt to different focal length characteristics difference.

Claims

1. A device for collaborative target tracking using both telephoto and focal length lenses, characterized in that, It includes a dual-optical-path system unit, a synchronous imaging unit, a tracking decision unit, and a tracking execution unit; The dual-optical-path system unit uses a beam splitter to split the incident light into two paths, one long focal length path and the other short focal length path, or it adopts a design with separate long focal length and short focal length lenses. The parallel optical axes of the lenses reduce the error in the center direction of the field of view, and the optical focus alignment controller ensures that the center field of view deviation of the two paths is less than 3 pixels @ 4K resolution. The synchronous imaging unit includes dual sensors and an image processor. The dual sensors are either visible light CMOS with global shutter, infrared sensors, or sensors with different wavelengths. The synchronous processing of dual RAW data is triggered by a single FPGA hardware chip. The image processor uses the RAW data from the dual sensors for ISP processing. The tracking decision unit determines the selection of tracking video data and the transition and switching of tracking algorithms based on the target size information; The tracking execution unit is used to execute tracking algorithms, applying corresponding tracking execution algorithms to sensor data from telephoto and short-focal-length optical paths.

2. A method for collaborative target tracking using telephoto and focal length lenses with the apparatus of claim 1, characterized in that, Specifically, this includes the following aspects: I. Cooperative Tracking Algorithm For long-focus paths, KCF correlation filter tracking is selected; for short-focus paths, SiamRPN tracking is selected. First, phase features are extracted using the telephoto path to resist motion blur. ； Then, texture and color features are extracted using the short focal path. ； Finally, cross-scale feature fusion Where α = 0.7 to 0.9 are empirical values; II. Switching control based on target size The dimensional characteristics are described using three key indicators: Pixel coverage is the target pixel area divided by the pixel area of ​​the image frame: The boundary distance ratio is the ratio of the distance from the target to the left boundary of the image to the distance to the right boundary of the image, and the ratio of the distance from the target to the top boundary of the image to the distance to the bottom boundary of the image. The minimum value of these two ratios is: Aspect ratio anomaly is the absolute value of the difference between the aspect ratio of the target pixels and the true aspect ratio of the target: in , The actual physical dimensions; When C < 5% or B > 0.3, the target size is extremely small, the lens uses a telephoto path, and a phase KCF correlation filter tracking strategy is adopted. When 5%≤C≤30% and B≥0.15, the target size is medium, the lens uses a telephoto path, and a feature fusion tracking strategy is adopted. When 25%≤C≤35%, the target size is a critical target, and the lens uses both telephoto and short-focus paths simultaneously, adopting an overlapping tracking strategy. When C > 30% or B < 0.15, the target size is large, the lens uses a short focal length path, and a local feature tracking strategy is adopted. When R > 1.5 and C > 20%, the target size is distorted, the lens uses a short focal length path, and a forced switching tracking strategy is adopted; III. The Decision-Making Process for Switching Between Telephoto and Short-Focal Lenses 1) As the aircraft gets closer to the target, it switches from telephoto to short-focus, which can be divided into four stages: Phase 1: Switching Triggers The process of switching triggers is as follows: (1) The long-focus tracker enable state triggers the decision layer; (2) The decision-making level triggers the short-focus system to start pre-tracking; (3) The short focal length system triggers the coordinate converter and requests the target coordinates H_matrix; (4) The coordinate converter converts the target coordinates and returns the homography matrix to adapt to the short focal length system; (5) The short focal length system triggers the target positioning function, that is, it enters the overlapping tracking stage of long focal length tracking and short focal length tracking; Phase Two: Overlap Tracking Period With a video frame rate of 60fps and an overlap tracker of 300ms, we will now track each frame of the image. (1) Dual-path parallel tracking, calculate the tracking frame for long-focus tracking and short-focus initialization. Telephoto Tracking Frame Short focal length initialization frame (2) Confidence fusion Long-range confidence Short focal length confidence ; (3) Weighted output Value parameters The value gradually changes from 0 to 1; Long focal length weight Short focal length weight (4) Anomaly detection Detect whether the tracking box features are abnormal If an anomaly is detected, the target should be recalibrated. ; Phase 3: Switching Completion Determination The switchover will be completed if any of the following conditions are met: (1) Short-focus confidence score > 0.85 for 5 consecutive frames; (2) The weighting coefficient alpha > 0.7; (3) Feature degradation occurs in telephoto lenses Phase Four: Release of Long-Focus Telephoto Resources (1) Retain the long-focus detector and operate at a low frequency of 1Hz; (2) Store target feature templates; (3) Release tracker resources; 2) Special cases of size variation An emergency switch is triggered when a target suddenly approaches; there are two types of triggering conditions: one is when the distance to the telephoto edge is greater than the target's distance. Second, aspect ratio anomaly. The solution is to immediately initiate short-focus tracking and switch directly, skipping the overlap period. 3) Adaptive optimization of control parameters (1) Dynamic threshold adjustment in, It is the target's motion speed (pixels / frame). It is the target acceleration. , These are all empirical coefficients; (2) Scenario-based hysteresis control IV. Ensuring Continuity of Tracking The measures to ensure continuous performance tracking are divided into pre-handover diagnostics and post-handover verification, as detailed below: (1) Pre-switching diagnosis Before switching, pre-switching diagnostic measures are added. First, it is determined whether the target size has reached the threshold. Second, the complexity of the target background is analyzed. If the target size and background are simple, the switching is started immediately. Otherwise, the stability of the target features is further determined. If the features are stable, the switching is performed normally. If the features are unstable, the switching is delayed and the target features are enhanced. (2) Verification after switching Geometric consistency check: Feature similarity verification: Motion continuity detection: .

Images