Perimeter intrusion continuous determination method based on multi-camera space-time correlation

Abstract

The present application relates to a perimeter intrusion continuous determination method based on multi-camera space-time correlation, acquires video image sequences of multiple cameras in a perimeter area, identifies target objects and forms corresponding local action clips; detects unfinished boundary actions of the target objects in the local action clips of the first camera, records corresponding perimeter positions and interruption states; searches candidate target objects in the perimeter positions and a preset time range in subsequent cameras, extracts starting action clips of the candidate target objects; conducts connection judgment according to the position continuity and action continuity of the unfinished boundary actions and the starting action clips; generates a cross-camera continuous action chain when the connection is established, and determines whether the target objects constitute a perimeter intrusion continuous event according to the continuous action chain. Through the cross-view connection mechanism, the local action clips are associated, the multiple verification and correction mechanisms are combined, the continuous tracking of the boundary-crossing behavior is realized, and the false alarm is effectively reduced, and the perimeter intrusion detection accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a method for determining continuous perimeter intrusion based on multi-camera spatiotemporal correlation. Background Technology

[0002] Current technologies, such as the multi-camera multi-target tracking method, device, storage medium, and electronic device disclosed in CN117670939A, mainly focus on trajectory-level association and stitching. The core idea is to merge local trajectories from different cameras through coordinate unification, trajectory similarity calculation, and threshold matching to form a global trajectory. In this paper, local trajectories are mapped to a unified coordinate system and trajectory similarity is calculated based on the shape similarity of the target boxes and the area intersection-union ratio. When the threshold condition is met, trajectory matching is completed and the global trajectory is updated. This method improves the success rate of cross-camera target matching to a certain extent, but its essence still remains at the spatial trajectory level of association. It lacks semantic understanding and phased modeling capabilities for the target behavior process, resulting in significant shortcomings in complex security scenarios, especially in continuous determination of perimeter intrusion. This technology typically uses complete trajectories or trajectory segments as matching units, lacking the recognition and utilization of incomplete behaviors. Once a target is obstructed in a camera's view, leaves the field of view, or its trajectory is interrupted due to a change in perspective, the system often has to rely on subsequent overall trajectory matching for compensation, rather than performing fine-grained matching based on the action phase. This is particularly prominent when the target performs actions that cross the perimeter and have obvious phases, which can easily lead to behavior fragmentation or even misjudgment.

[0003] This method primarily calculates trajectory similarity based on geometric and positional features, with limited consideration for action semantics and continuity. It fails to distinguish between a target performing a continuous process of the same action and new actions occurring independently in different cameras. Particularly when complex behaviors such as loitering, probing, and repeated approaching occur near the perimeter, it easily misjudges independent actions as continuous intrusions, or vice versa, thus affecting accuracy. While the document mitigates trajectory breakage through prediction, this prediction relies heavily on positional continuity and movement trend inference, lacking constraints on action phases. Therefore, the prediction reliability is low when actions change or transitions occur, failing to effectively support the judgment of continuous intrusion behavior. This technology typically only focuses on whether the target enters a specific area or meets certain trajectory conditions. The method lacks a mechanism to determine the degree of boundary crossing completion, meaning it cannot distinguish whether the target has completed a valid boundary crossing or is merely making a tentative entry. For example, a short-term entry followed by a rapid return may still be classified as an intrusion event in existing methods, leading to a high false alarm rate. This method largely relies on overlapping areas of camera fields of view for trajectory matching. In scenarios with non-complete or weak overlap, the reliability of cross-camera associations significantly decreases. In actual perimeter deployments, there are often gaps or non-ideal overlapping areas between cameras, making it difficult to guarantee continuity when relying solely on trajectory similarity for association. Furthermore, this method generally lacks a mechanism to handle reverse behavior changes. When a target retreats or returns to the outside after entering the perimeter, the system typically cannot promptly identify the impact of this behavior on the nature of the event and may maintain the original intrusion judgment result, lacking dynamic correction capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide a method for determining the continuity of perimeter intrusion based on the spatiotemporal correlation of multiple cameras, thereby solving some of the drawbacks and shortcomings pointed out in the background art.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows: a method for continuous determination of perimeter intrusion based on multi-camera spatiotemporal correlation, comprising: acquiring video image sequences of multiple cameras in the perimeter area, identifying target objects and forming corresponding local action segments;

[0006] In the local motion segment of the first camera, the incomplete boundary action of the target object is detected, and the corresponding perimeter position and interruption state are recorded; in the subsequent cameras, candidate target objects are searched within the perimeter position and a preset time range, and the starting motion segment of the candidate target object is extracted.

[0007] The continuity of the position and the continuity of the action between the incomplete boundary action and the starting action segment are used to make a connection judgment; when the connection is established, a continuous action chain across cameras is generated, and the target object is determined to constitute a continuous perimeter intrusion event based on the continuous action chain.

[0008] Furthermore, when recording the perimeter position and interruption state, the action interruption point and action interruption location of the incomplete boundary action are recorded; the candidate target object in the subsequent camera is only retained when the starting action corresponds to the action interruption location.

[0009] Furthermore, the acceptance determination includes: determining whether the starting action of the candidate target object is a successor action of the incomplete boundary action, and determining whether the starting action meets the determination condition of an independent starting boundary action; acceptance is determined to be successful only when the starting action is a successor action and does not meet the determination condition of an independent starting boundary action.

[0010] Furthermore, when determining a continuous perimeter intrusion event based on the continuous action chain, it continues to detect whether a reverse correction action segment occurs; when the continuous action chain reaches the inner perimeter and no reverse correction action segment is detected, it is determined that a continuous perimeter intrusion event has occurred; when the reverse correction action segment is detected, the determination is terminated.

[0011] Furthermore, the termination point of the incomplete boundary action is recorded, and the starting action is matched with the termination point; only when the starting action can complete the corresponding incomplete boundary passage process is it determined to be a subsequent action.

[0012] Furthermore, it is detected whether the candidate target object has a boundary approach action or a boundary probing action before the starting action; when there is no boundary approach action or boundary probing action, and the starting action corresponds to an intermediate action in the boundary crossing process, it is determined that the starting action does not meet the determination condition of an independent starting boundary action.

[0013] Furthermore, the reverse correction action segment is an action segment that moves back towards the perimeter or returns to the outside of the perimeter after the target object reaches the inner side of the continuous action chain; when the reverse correction action segment is opposite to the boundary crossing direction of the continuous action chain, it is determined that the nature of the continuous action chain has changed.

[0014] Furthermore, after the continuous action chain reaches the inner perimeter, the reverse correction action segment is continuously detected within a preset verification period; only when the target object remains active within the perimeter during the preset verification period and the reverse correction action segment is not detected is it determined that a continuous perimeter intrusion event has been constituted.

[0015] Furthermore, when the reverse correction action segment is detected, the continuous action chain is rolled back to the most recent boundary contact action segment, and the corresponding event is updated to an incomplete boundary crossing event.

[0016] Furthermore, the preset verification period starts from the moment the continuous action chain first reaches the inner side of the perimeter; if the target object moves away from the perimeter or stays inside the perimeter during the preset verification period, it is determined that it is maintaining the activity state inside the perimeter; if it stays in the preset adjacent area inside the perimeter for no more than a preset threshold and then returns to the perimeter position, it is determined that it is not maintaining the activity state inside the perimeter.

[0017] The beneficial effects of this invention are as follows: By introducing a cross-view continuity mechanism for incomplete boundary actions in multi-camera scenarios, the local action segments of the target object in different cameras are spatiotemporally correlated. This effectively solves the action fragmentation problem existing in traditional single-camera or simple cross-camera switching, enabling continuous tracking and overall reconstruction of the target object's boundary-crossing behavior. By recording the action interruption point and the interruption location, and combining it with the initial action segment for continuation matching and subsequent action determination, the accuracy of cross-camera target matching can be significantly improved, reducing mismatch problems caused by target occlusion, viewpoint switching, or short-term loss, thereby improving the stability and reliability of perimeter intrusion detection.

[0018] By introducing an independent starting boundary action exclusion mechanism and a reverse correction action detection mechanism, boundary crossing behavior is dynamically verified based on the construction of a continuous action chain. This not only avoids misjudging independent boundary behaviors as continuous intrusions but also identifies retreat or probing behaviors of the target object after crossing the boundary, and corrects the event status to an incomplete boundary crossing event through the rollback mechanism. Furthermore, by setting a preset verification period and combining it with the target object's activity status inside the perimeter, the system can effectively distinguish between brief boundary crossings and genuine intrusion behaviors, thereby significantly reducing the false alarm rate and improving the accuracy and practicality of determining continuous perimeter intrusion events. Attached Figure Description

[0019] Figure 1 This is a functional relationship diagram of the perimeter intrusion continuous determination based on multi-camera spatiotemporal correlation in this invention.

[0020] Figure 2 This is a scatter plot for classifying candidate objects in Embodiment 1 of the present invention.

[0021] Figure 3 This is a diagram showing the generation of a local action segment timeline and a continuous action chain in Embodiment 1 of the present invention.

[0022] Figure 4 This is a probability determination diagram for continuous scoring and exclusion of independent starting boundary actions in Embodiment 1 of the present invention.

[0023] Figure 5 This is a timing diagram of the verification period and event rollback of the cross-camera continuous action chain in Embodiment 2 of the present invention.

[0024] Figure 6This is a step diagram of the inner activity retention state in Embodiment 2 of the present invention.

[0025] Figure 7 This is a bubble diagram of the reverse correction action decision in Embodiment 2 of the present invention. Detailed Implementation

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] Combined with appendix Figure 1 This invention is a perimeter intrusion continuity determination method based on multi-camera spatiotemporal correlation. Multiple cameras are deployed in the perimeter area to form continuous coverage of the same perimeter. Each camera acquires video image sequences according to a unified time reference and transmits them to the processing unit. The video image sequences are analyzed frame by frame. Target objects entering the monitoring range are identified through target detection and target tracking methods. The behavior of the target objects is segmented based on the spatial position and posture changes between consecutive frames. Image segments with continuous action characteristics are divided into local action segments and corresponding temporal sequence relationships are established.

[0028] In the local motion segment corresponding to the first camera, the analysis is performed on whether the target object performs boundary-related actions. When the target object is detected to be approaching the perimeter and starting to perform an action to cross or touch the boundary but not completing the crossing process, the action is identified as an incomplete boundary action. At the same time, the perimeter spatial position where the action occurs and the action interruption status are recorded. The interruption status is used to characterize the situation where the target object stops or leaves the current camera's field of view during the action execution.

[0029] For the video image sequence acquired by the subsequent camera, a search for candidate target objects is conducted in the spatially adjacent area corresponding to the recorded perimeter position and within a preset time range. Target objects corresponding to the aforementioned incomplete boundary actions are selected through position matching and time constraints. Action analysis is performed on the selected candidate target objects, and the action segments of the initial stage are extracted from the continuous frames after they enter the screen as the starting action segments for subsequent cross-camera action inheritance and continuity determination.

[0030] First, based on the perimeter spatial location, the termination position of the incomplete boundary action is matched with the starting position of the initial action segment to determine whether they are within the same boundary area or a spatially adjacent range, thus establishing positional continuity. Then, considering the temporal sequence, it is confirmed whether the time of the initial action segment falls within a preset time range, thereby eliminating obviously discontinuous target objects.

[0031] Based on the requirement of positional continuity, the continuity of action is determined. By analyzing the action stages of the incomplete boundary action and the corresponding interruption points, combined with the action type and direction exhibited by the initial action segment, it is determined whether the initial action can serve as a subsequent continuation of the incomplete boundary action. If the initial action and the incomplete boundary action can form a sequential connection in terms of action stages, and there is no independently initiated boundary approach or probing behavior, then the succession relationship between the two is determined to be valid.

[0032] When the connection is established, local motion clips from different cameras are correlated and spliced ​​in chronological order to construct a cross-camera continuous motion chain. This continuous motion chain represents the complete behavioral trajectory of the target object from different camera perspectives. Based on the continuous motion chain, a comprehensive analysis is performed to determine whether the target object has completed a continuous passage from the outside to the inside of the perimeter. When the continuous motion chain covers the complete boundary passage path without any interruption or reverse behavior, the target object is determined to constitute a continuous perimeter intrusion event, thereby improving the accuracy and continuity of perimeter intrusion determination in cross-camera scenarios.

[0033] When recording the perimeter position and interruption status, the incomplete boundary actions are further annotated. In addition to recording the spatial position of the target object at the perimeter, the interruption point and the interruption location are also recorded. The interruption point characterizes the interruption frame position of the target object in the time series, while the interruption location characterizes the state of the body parts involved in the target object's boundary crossing process, such as being in the initial or intermediate stage of crossing. By jointly recording the interruption point and interruption location, a structured description of the incomplete boundary actions can be formed.

[0034] When screening candidate target objects in the video image sequence of subsequent cameras, the interruption point of the action is used as a key constraint condition. The starting action of the candidate target object is compared. Only when the action stage reflected by the starting action is consistent with the interruption point of the action or can form a connection relationship is the candidate target object retained, thereby eliminating interference targets that are unrelated to the aforementioned incomplete boundary action.

[0035] During the judgment process, the successor relationship analysis is performed on the initial action of the candidate target object. By comparing the termination stage of the incomplete boundary action with the initial stage of the initial action, it is determined whether the initial action can be regarded as a continuation of the aforementioned action, thus determining whether it belongs to the successor action. At the same time, it is analyzed whether the initial action meets the judgment condition of an independent starting boundary action. By detecting whether there is boundary approach behavior or boundary probing behavior before the initial action, and whether the initial action is in the initial stage of a complete boundary crossing process, it is determined whether it has independent starting characteristics.

[0036] Only when the starting action is determined to be a follow-up action of an incomplete boundary action and does not have the characteristics of an independent starting boundary action, is it determined that there is a valid succession relationship between the candidate target object and the aforementioned incomplete boundary action, thereby providing a reliable basis for the construction of a continuous action chain across cameras.

[0037] When identifying incomplete boundary actions, the action state of the target object at the moment of action termination is further annotated, and the termination action part is recorded to reflect its specific stage in the boundary crossing process. The termination action part is used to describe the positional relationship and action form of the key parts of the target object that participated in boundary crossing when interrupted, thereby determining its incomplete crossing process.

[0038] After extracting the starting action segment of the candidate target object from subsequent cameras, the starting action is matched with the ending action segment. By comparing the correspondence between the two in terms of action stages and the consistency of action direction, it is determined whether the starting action can continue the unfinished boundary crossing process. If the starting action can logically continue the unfinished stage corresponding to the ending action segment and continue to complete the remaining boundary crossing path, then the starting action is determined to be a subsequent action; if the starting action cannot form a continuous action continuation relationship, it is not considered a subsequent action.

[0039] When determining whether a starting action possesses independent initiation characteristics, a backtracking analysis is performed on the behavior of the candidate target object prior to the starting action to detect whether it exhibits boundary approach or boundary probing actions. Boundary approach actions characterize the behavior of a target object gradually approaching the perimeter, while boundary probing actions characterize the behavior of a target object touching or attempting to cross the boundary. When the detection results indicate that none of the above behaviors existed before the starting action, and the starting action itself corresponds to an intermediate stage rather than the initial stage in the boundary crossing process, the starting action is determined not to meet the conditions for an independent initiation boundary action. This is used to exclude independently occurring boundary crossing behaviors and enhance the accuracy of cross-camera action transitions.

[0040] A continuous motion chain integrates local motion clips from multiple cameras in chronological order to represent the continuous movement of a target object from the outside to the inside of the perimeter. During the judgment process, the actions at each stage of the continuous motion chain are continuously analyzed to confirm whether the target object has completed a full boundary crossing path.

[0041] In the determination of continuous action chains, a detection mechanism for reverse correction action segments is introduced. Reverse correction action segments characterize the behavior of a target object moving back towards the perimeter or returning to the outside of the perimeter after completing or about to complete a boundary crossing. By analyzing the target object's movement direction and displacement trend near the perimeter, when its movement direction is detected to be opposite to the original boundary crossing direction, the corresponding action is identified as a reverse correction action segment.

[0042] After the continuous action chain reaches the inner perimeter, the subsequent behavior of the target object continues to be tracked and analyzed. If no reverse correction action segment is detected at this stage, and the target object continues to move inward or operate in the inner region, the continuous action chain is determined to constitute a perimeter intrusion continuous event, thus confirming that the target object has completed a valid boundary crossing. Conversely, if a reverse correction action segment is detected during the determination of the continuous action chain, it indicates that the target object has not formed a stable boundary crossing result. At this time, the determination of the continuous action chain as a perimeter intrusion continuous event is stopped to avoid misjudgment and improve the accuracy of the overall determination.

[0043] Reverse correction action segments describe behavioral changes of a target object after a continuous action chain has reached the inner perimeter boundary. This behavior manifests as the target object changing its movement from the original boundary-crossing direction to moving towards the perimeter, or directly returning to the outer perimeter area. By analyzing the displacement direction and positional changes of the target object in consecutive frames, when its movement trend changes from extending inwards to approaching or crossing the perimeter, the corresponding action segment is identified as a reverse correction action segment. When the movement direction of the reverse correction action segment is opposite to the boundary-crossing direction represented by the continuous action chain, it is determined that the continuous action chain no longer maintains its original boundary-crossing attribute, and therefore it is considered to lack the basis for determining a continuous perimeter intrusion event.

[0044] After the continuous action chain first reaches the inner perimeter, a preset verification period is set to continuously monitor the subsequent behavior of the target object. During this verification period, the movement state of the target object is continuously tracked, with a focus on detecting whether reverse correction action segments occur. Simultaneously, the activity of the target object inside the perimeter is analyzed. If the target object remains inside the perimeter and shows a trend of moving away from the perimeter, or remains active in the inner region, and no reverse correction action segments are detected throughout the verification period, its behavior is determined to constitute a continuous perimeter intrusion event, thus confirming the stability of the boundary crossing behavior.

[0045] When a reverse correction action segment is detected during the verification process, the established continuous action chain is rolled back to the time node corresponding to the most recent boundary contact action segment, and the current event state is updated to an incomplete outbound event.

[0046] The preset verification period is used to confirm the stability of the target object's behavior after it crosses the boundary. Its start time is set to the moment when the continuous action chain first reaches the inner edge of the boundary. At this moment, the system marks the position of the target object and uses this time point as the timing start point to continuously track the target object's motion trajectory and stationary state in subsequent consecutive frames.

[0047] During a preset verification period, the positional changes and direction of movement of the target object relative to the perimeter are analyzed to determine whether it remains in an activity state inside the perimeter. When the target object exhibits inward displacement away from the perimeter during this period, i.e., its direction of movement continuously points towards the inner area of ​​the perimeter, or it continuously stays within the inner area of ​​the perimeter without showing obvious return movement towards the perimeter, it is determined that it remains in an activity state inside the perimeter, thus indicating that the boundary crossing behavior is stable.

[0048] To distinguish between brief entry and valid boundary crossing, a preset range is set in the adjacent area inside the perimeter, and the time the target object stays in this area is statistically analyzed. When the target object enters this adjacent area and stays for no more than a preset threshold and then returns to the perimeter position, it is determined that it has not maintained an activity state inside the perimeter, indicating that its behavior is a tentative entry or a short-term boundary crossing, and does not constitute a stable intrusion.

[0049] Example 1:

[0050] Three cameras are deployed along the boundary of a certain perimeter area, covering the adjacent outer approach area, the boundary crossing area, and the inner landing area, respectively. There are partial field-of-view overlaps between adjacent cameras, but the first and third cameras do not directly overlap. The system continuously acquires video image sequences from the three cameras, detects, tracks, and recognizes the target objects in the images, and segments them into local action segments according to action semantics. In this embodiment, the target objects correspond to segments F1 to F4 in the first camera and segments G1 to G5 in the second camera.

[0051] In the image sequence from the first camera, the system detected a target object approaching the boundary from the outside, exhibiting a series of actions including deceleration, upper limb elevation, grasping the upper edge of the boundary, and body lifting. Segment F1, with a length of 18 frames, represents the approaching action from the outside; segment F2, with a length of 22 frames, represents the boundary probing action; segment F3, with a length of 31 frames, represents the grasping and lifting action; and segment F4, with a length of 14 frames, represents the incomplete boundary crossing action. The system detected image occlusion in frame 11 of F4. The target object's chest and above had already crossed the upper edge of the boundary, while the pelvis and lower limbs remained outside the boundary; therefore, F4 was classified as an incomplete boundary crossing action.

[0052] For this incomplete boundary action, the system records the corresponding boundary position as boundary segment K7, with a normalized coordinate of 0.62. The recorded interruption status is occlusion interruption, the recorded action interruption point is frame 11 of F4, the recorded action interruption location is the pelvis and right lower limb failing to cross the boundary, and the recorded termination action location is the starting point of the right leg crossing. Since this termination action location indicates that the target object has completed upper limb support and entered the intermediate stage of the boundary crossing process, subsequent candidate target objects in the camera must be able to start with the right leg continuing to step down or the pelvis continuing to move forward in order to enter the retention set.

[0053] The system then searched for candidate targets in the second camera, using the mapped position of segment K7 and a preset time range of 2.4 seconds as constraints, and detected a total of 3 candidate targets. The first candidate target's starting segment G1 appeared as a leftward hovering motion followed by a movement towards the boundary, with the initial action being a tentative hand movement, which did not correspond to the interruption point, so it was eliminated. The second candidate target's starting segment G2 showed an independent jump after a natural swing of both arms. Although the position was close to K7, the starting point was a forward swing of the upper limbs, which did not correspond to the right leg crossing the starting point, so it was also eliminated. The third candidate target's starting segment G3 was 19 frames long. Its first 6 frames showed a forward pelvic movement and a continued downward and inward swing of the right leg, which was consistent with the interruption point and the termination point recorded by the first camera, so it was retained as the only candidate target. Figure 2 The results indicate that the system detected three candidate objects in the second camera, constrained by the mapping position of segment K7 and a preset time range of 2.4 seconds. Candidate object 1 was eliminated because its initial movement was a hand probing motion, which did not correspond to the interruption point. Candidate object 2 was eliminated because its initial movement was an upper limb forward swing, which did not correspond to the termination point. Only candidate object 3, exhibiting a pelvic forward thrust and a continued downward swing of the right leg, matched the interruption and termination points recorded by the first camera, and was therefore retained. This result demonstrates that the system does not simply rely on proximity for selection, but rather achieves effective filtering through the complementary relationship between the interruption point, termination point, and initial movement point.

[0054] For the object being retained, the system first determines whether its initial movement is a follow-up movement to an incomplete boundary movement. The system then matches the starting point of G3 with the ending point of F4. When F4 stops, the upper limb support is complete, the pelvis presses against the edge, and the right leg begins to cross the boundary. When G3 starts, the pelvis continues to move forward, the right leg swings down, and the center of gravity shifts inward. Since G3 does not involve a renewed grab, probe, or initial jump, but rather directly follows the mid-stage movement of the boundary crossing process, the system determines that G3 can complete the corresponding incomplete boundary crossing process, thus satisfying the follow-up movement determination condition.

[0055] After the connection is established, the system splices segments F3 and F4 from the first camera with segments G3, G4, and G5 from the second camera to generate a continuous action chain across cameras. G4 represents the target object's center of gravity shifting to the inside of the perimeter and completing the lower limb landing, with a length of 16 frames. G5 represents the target object continuing to move forward inside the perimeter, with a length of 21 frames. Because the continuous action chain begins with the grabbing and lifting action in the first camera, proceeds through the incomplete crossing action, and is then completed by the pelvic forward movement, right leg swing, and inside landing actions in the second camera, the entire boundary passage process maintains spatiotemporal semantic unity and continuity. Figure 3 The corresponding data shows that the segment lengths in the first camera are F1=18 frames, F2=22 frames, F3=31 frames, and F4=14 frames, respectively, while the key segment lengths in the second camera are G3=19 frames, G4=16 frames, and G5=21 frames, respectively. Among them, F4 is interrupted at the 11th frame, while the first segment of G3 directly shows the pelvis moving forward and the right leg continuing to press down and swing inward, without any re-grasping, probing, or jumping actions. Therefore, G3 is a continuation of the boundary crossing process that was not completed in F4. Thus, the system splices F3, F4 and G3, G4, G5 into a continuous action chain across cameras, indicating that the entire boundary crossing process maintains a singleness and continuity in spatiotemporal semantics.

[0056] To improve the objectivity of the continuity judgment, the system further calculates a comprehensive score for positional continuity and action continuity. The following formula is used:

[0057]

[0058] in, Indicates the deviation of the perimeter position. Indicates deviation in the action phase. Indicates the matching coefficient for the action part. , , As the weight. In this embodiment, take =0.35, =0.40, =0.25, positional deviation Location scale parameters Action phase deviation Stage scale parameters Complementary matching coefficient Substituting, we get:

[0059]

[0060] The system sets the acceptance threshold to 0.70 because:

[0061]

[0062] This indicates that the candidate object has a high degree of continuity with the incomplete boundary action in the previous camera in terms of both perimeter position and action phase, and the succession relationship is established.

[0063] After the connection is established, the system still needs to determine whether G3 meets the criteria for an independent starting boundary action. To do this, the system backtracks the 20 frames preceding G3 from the second camera to detect any boundary approach or boundary probing actions. The detection results show that no continuous displacement from the outside towards the boundary occurred before G3, indicating insufficient evidence of boundary approach actions. There was no evidence of repeated hand contact with the edge, testing the height, or leaning forward to observe; the evidence of boundary probing actions was minimal. Meanwhile, the initial movement of G3 is directly manifested as pelvic forward movement and right leg swing, which is clearly an intermediate movement in the boundary passage process. Therefore, the quantification value of the intermediate movement is... .

[0064] The system uses the following formula to calculate the probability of excluding independent initial boundary actions:

[0065]

[0066] in, This indicates the amount of evidence related to the boundary approaching the action. This indicates the amount of evidence for boundary probing actions. This is a quantized value indicating whether an action is intermediate. In this embodiment, it is taken as... , , , Substituting, we get:

[0067]

[0068] The system sets the threshold for establishing an independent starting boundary action to 0.50, because:

[0069]

[0070] This indicates that the initial action does not meet the criteria for an independent initial boundary action. Figure 4 The corresponding result is obtained after calculation based on the parameters in the embodiment. Since this value is greater than the acceptance threshold of 0.70, it indicates that candidate object 3 meets the acceptance conditions in terms of positional continuity, action phase continuity, and action part completion matching; at the same time, the calculated result is... Since this value is less than the threshold of 0.50 for independent starting boundary action, it indicates that there are no valid boundary approach actions or boundary probing actions before G3, and G3 itself is an intermediate action in the boundary crossing process.

[0071] Based on this, the system arrives at the following judgment: First, G3 can complete the incomplete boundary crossing process corresponding to F4, and is a subsequent action. Second, before G3 appears, there are no valid boundary approach actions or boundary exploration actions, and G3 itself is an intermediate action in the boundary crossing process, thus not satisfying the condition of an independent starting boundary action. Therefore, this candidate object simultaneously satisfies both the constraint of being a subsequent action and not satisfying the condition of an independent starting boundary action, and the system determines that the inheritance is successful.

[0072] Example 2:

[0073] A front camera, a rear camera, and an inner supplementary camera are deployed along the boundary extension direction of a perimeter. The front camera mainly covers the area approaching the outer edge of the perimeter and the boundary contact area. The rear camera mainly covers the inner adjacent area after the boundary is crossed. The inner supplementary camera covers a deeper layer of activity area inside the perimeter. The system continuously acquires video image sequences from each camera, detects target objects, performs cross-camera tracking and action segmentation, and forms local action segments according to action semantics. In this embodiment, the target object forms segments A1, A2, and A3 in the front camera, segments B1, B2, B3, and B4 in the rear camera, and segment C1 in the inner supplementary camera.

[0074] In the previous camera, the target object first approaches the boundary from the outside of the perimeter. Segment A1, with a length of 20 frames, represents the approaching action; segment A2, with a length of 28 frames, represents the action of grasping the boundary and lifting up; and segment A3, with a length of 17 frames, represents the action of pressing down after the center of gravity crosses the upper edge of the boundary. Since the main body of the target object has crossed the boundary center line in the last segment of A3, and a continuing action occurs at the corresponding position in the next camera, the system, based on position association and action continuity, splices A2, A3, and B1, B2 in the next camera into a cross-camera continuous action chain L1. Among them, B1, with a length of 15 frames, represents the lower limbs swinging down and landing on the inside; and B2, with a length of 18 frames, represents the body completely entering the inside of the perimeter and moving away from the boundary line by 1.6m.

[0075] The system first confirms that the target object has reached the inner perimeter boundary in frame 9 of B2, and records the corresponding moment of that frame as the verification start point. From that moment, the system initiates a preset verification period, with a duration of 4.0 seconds, to continuously detect the existence of reverse correction action segments. Reverse correction action segments refer to actions in which the target object, after reaching the inner perimeter boundary in a continuous action chain, again moves back towards the perimeter boundary or returns to the outer perimeter boundary. The system does not make a final judgment immediately upon the target object's first arrival at the inner perimeter boundary, but instead continues to combine the footage from the next camera and the inner supplementary camera to determine the nature of its subsequent activities. Figure 5The corresponding information indicates that the system splices A2 and A3 from the previous camera with B1 and B2 from the next camera into a continuous action chain L1. The system first confirms that the target object has reached the inner edge of the perimeter in the 9th frame of B2, and records the time corresponding to that frame as the verification start point. A preset verification period of 4.0s is started from that time. After detecting the back movement represented by B4 and C1 in the second half of the verification period, the system determines that the nature of the event has changed, and reverts the continuous action chain to the most recent boundary contact segment D0, canceling the intrusion confirmation state formed after B2.

[0076] In the first half of the verification period, the target object first forms segment B3 in the adjacent area inside the perimeter. B3 is 22 frames long and consists of a short lateral movement and a stop within the perimeter. System statistics show that after initially entering the inner perimeter, the target object continues to move 1.2m away from the boundary, and then stops for 1.1s in the inner region at a distance of 1.4m to 1.6m from the boundary. Since this displacement direction is away from the boundary, and the stop position is in the inner activity area rather than the edge-hugging area, the system initially classifies this phase as maintaining an activity state inside the perimeter.

[0077] To quantitatively evaluate whether the target object maintains its activity state inside the perimeter during the verification period, the system uses the inside activity retention index, the formula of which is as follows:

[0078]

[0079] in, For the preset verification period, Indicates the target object at time An indicator function for whether the activity state inside the perimeter is maintained. In this embodiment, the system discretizes the verification period into 8 equal-length statistical units, each statistical unit being 0.5s. Within the first 5 statistical units, the target object is located inside the perimeter and exhibits displacement away from the boundary or stable residence; therefore, the value is taken as... =1. In the 6th statistical unit, the target object begins to turn towards the boundary, but has not yet entered the edge-hugging return state. =0.5. In the 7th and 8th statistical units, the target object continued to move back towards the boundary and re-entered the inner adjacent area 0.4m from the boundary. =0. Therefore, we can approximate the following:

[0080]

[0081] The system sets the inner activity retention threshold to 0.75. Because... If the value is below the threshold, it indicates that the target object did not maintain a continuous activity state inside the perimeter during the entire verification period, and further determination is needed to determine whether there are reverse correction action segments. Figure 6The corresponding data indicates that the system discretizes the 4.0s verification period into 8 equal-length statistical units, each unit being 0.5s. The first 5 statistical units have a value of 1, the 6th statistical unit has a value of 0.5, and the 7th and 8th statistical units have a value of 0. Based on this, the medial activity retention rate is calculated. The result is below the retention threshold of 0.75. This indicates that the target object did not maintain a continuous activity state inside the perimeter throughout the verification period, and further judgment needs to be made by combining the results of the reverse correction direction.

[0082] In the latter half of the verification period, the next camera detected segment B4, and the inner supplementary camera simultaneously detected segment C1. B4, with a length of 19 frames, depicts the target object gradually moving 1.1 meters back towards the boundary from a position 1.5 meters away. C1, with a length of 14 frames, depicts the target object pausing for 0.6 seconds after moving back to near the boundary, and then exhibiting a turning motion towards the outside of the perimeter. The system denotes the main motion vector extracted in this phase as... The horizontal component points to the boundary, and the preset cross-boundary direction vector is denoted as... , indicates the direction from the outside of the perimeter to the inside of the perimeter.

[0083] To determine whether the return movement is opposite to the original out-of-bounds direction, the system uses a reverse correction direction discrimination coefficient, the formula of which is expressed as:

[0084]

[0085] Substituting the above data, we get:

[0086]

[0087]

[0088]

[0089] Therefore:

[0090]

[0091] The system sets the opposite direction determination threshold to -0.30. Because... If the movement is significantly below the threshold, it indicates that the target object's movement direction at this stage is opposite to the original boundary crossing direction, which is a significant return movement towards the perimeter. Therefore, it should be determined that there is a reverse correction action segment. Figure 7 The corresponding data indicates that the system makes joint decisions based on the reverse correction direction discrimination coefficient, the dwell time in the neighboring area, and the return distance, where the main motion vector is the determining factor. Preset outbound direction vector Calculations yielded Since this value is significantly lower than the opposite direction determination threshold of -0.30, it indicates that the target object's movement direction in the latter half of the verification period is opposite to the original boundary crossing direction. Combined with its return distance of 1.1m, re-entering the inner adjacent area 0.4m from the boundary, and returning to the perimeter position after staying for 0.6s, it can be determined that there is a reverse correction action segment. The event nature of the continuous action chain L1 has changed and should be updated to an incomplete boundary crossing event.

[0092] Combining the retention rate results and direction discrimination results, the system further examines the dwelling characteristics of the target object in the adjacent area inside the perimeter. In this embodiment, the preset adjacent area inside the perimeter is defined as an area no more than 0.5m from the boundary, and the short-term dwell threshold in the adjacent area is set to 0.8s. The detection results show that after the target object re-enters this adjacent area at the end of the verification period, its dwell time is 0.6s, which does not exceed the dwell threshold, but it does not continue to move inward afterward, but returns to the perimeter position. According to the preset rules, if the target object stays in the adjacent area inside the perimeter for no more than the threshold and then returns to the perimeter position, it should be judged as not maintaining the activity state inside the perimeter.

[0093] Since the system has detected a reverse correction action segment, and this reverse correction action is opposite to the boundary crossing direction of the original continuous action chain, it determines that the event nature of continuous action chain L1 has changed and is no longer maintained as a confirmed intrusion result. At this time, the system rolls back the continuous action chain to the most recent boundary contact action segment according to the rollback rules. In this embodiment, the most recent boundary contact action segment is the cross-boundary contact segment D0 corresponding to the end of A3 in the previous camera and the beginning of B1 in the next camera. The system retains the approach, grab, and cross-attack information before D0, but cancels the intrusion confirmation state formed after B2.

[0094] After the rollback is complete, the system updates the original event status from "Pending Intrusion Event" to "Incomplete Boundary Crossing Event." In the updated event log, the endpoint of the continuous action chain is no longer located at the inner stable activity phase, but at the most recent effective boundary contact node, with an added reverse correction marker, a rollback distance of 1.1m, a dwell time in the adjacent area of ​​0.6s, and a direction discrimination coefficient. Since the target object once reached the inner perimeter, it significantly shifted back and returned to the boundary position during the verification period, thus failing to meet the final judgment condition of maintaining the inner perimeter activity state and not detecting any reverse correction action segments during the preset verification period.

Claims

1. A method for determining the continuity of perimeter intrusion based on multi-camera spatiotemporal correlation, characterized in that... include: Acquire video image sequences from multiple cameras in the perimeter area, identify target objects, and generate corresponding local action segments; In the local motion segment of the first camera, the incomplete boundary action of the target object is detected, and the corresponding perimeter position and interruption state are recorded; in the subsequent cameras, candidate target objects are searched within the perimeter position and a preset time range, and the starting motion segment of the candidate target object is extracted. The continuity of the position and the continuity of the action between the incomplete boundary action and the starting action segment are used to make a connection judgment; when the connection is established, a continuous action chain across cameras is generated, and the target object is determined to constitute a continuous perimeter intrusion event based on the continuous action chain.

2. The perimeter intrusion continuity determination method based on multi-camera spatiotemporal correlation according to claim 1, characterized in that, When recording the perimeter position and interruption status, the interruption point and interruption location of the incomplete boundary action are recorded; candidate target objects in subsequent cameras are only retained when the starting action corresponds to the interruption location.

3. The perimeter intrusion continuity determination method based on multi-camera spatiotemporal correlation according to claim 1, characterized in that, The acceptance determination includes: determining whether the starting action of the candidate target object is a follow-up action of the incomplete boundary action, and determining whether the starting action meets the determination condition of an independent starting boundary action; acceptance is determined to be successful only when the starting action is a follow-up action and does not meet the determination condition of an independent starting boundary action.

4. The perimeter intrusion continuity determination method based on multi-camera spatiotemporal correlation according to claim 1, characterized in that, When determining a continuous perimeter intrusion event based on the continuous action chain, the system continues to detect whether a reverse correction action segment occurs. When the continuous action chain reaches the inner perimeter and no reverse correction action segment is detected, a continuous perimeter intrusion event is determined to occur. When the reverse correction action segment is detected, the determination is terminated.

5. The perimeter intrusion continuity determination method based on multi-camera spatiotemporal correlation according to claim 3, characterized in that, Record the termination point of the incomplete boundary action, and perform a follow-up matching between the starting action and the termination point; only when the starting action can complete the corresponding incomplete boundary passage process is it determined to be a subsequent action.

6. The perimeter intrusion continuity determination method based on multi-camera spatiotemporal correlation according to claim 3, characterized in that, The system detects whether the candidate target object has a boundary approach action or a boundary probing action before the starting action; when there is no boundary approach action or boundary probing action, and the starting action corresponds to an intermediate action in the boundary crossing process, the system determines that the starting action does not meet the determination condition of an independent starting boundary action.

7. The perimeter intrusion continuity determination method based on multi-camera spatiotemporal correlation according to claim 4, characterized in that, The reverse correction action segment is an action segment in which the target object moves back towards the perimeter or returns to the outside of the perimeter after the continuous action chain reaches the inside of the perimeter; when the reverse correction action segment is opposite to the boundary crossing direction of the continuous action chain, it is determined that the nature of the continuous action chain has changed.

8. The perimeter intrusion continuity determination method based on multi-camera spatiotemporal correlation according to claim 4, characterized in that, After the continuous action chain reaches the inner perimeter, the reverse correction action segment is continuously detected within a preset verification period; only when the target object remains active inside the perimeter during the preset verification period and the reverse correction action segment is not detected is it determined to constitute a continuous perimeter intrusion event.

9. The perimeter intrusion continuity determination method based on multi-camera spatiotemporal correlation according to claim 4, characterized in that, When the reverse correction action segment is detected, the continuous action chain is rolled back to the most recent boundary contact action segment, and the corresponding event is updated to an incomplete boundary crossing event.

10. The perimeter intrusion continuity determination method based on multi-camera spatiotemporal correlation according to claim 8, characterized in that, The preset verification period starts from the moment the continuous action chain first reaches the inner side of the perimeter; if the target object moves away from the perimeter or stays inside the perimeter during the preset verification period, it is determined that it is maintaining the activity state inside the perimeter; if it stays in the preset adjacent area inside the perimeter for no more than a preset threshold and then returns to the perimeter position, it is determined that it is not maintaining the activity state inside the perimeter.

Images