A sensitive area automatic identification and shielding method and system for high-position camera deployment

Abstract

The application relates to the technical field of video monitoring privacy protection, and discloses a sensitive area automatic identification and shielding method and system for high-position camera deployment, wherein abnormal illumination events are detected by collecting window area illumination data, time domain features are extracted to classify abnormal event types, event semantics are inferred based on abnormal-semantic mapping rules, potential conflict targets in a tracking task are identified, monitoring priority weights and tracking channel opening authorization levels are differentially adjusted according to event semantic differences, a coordination scheme is calculated in combination with tracking priority, a semantic perception shielding mask is generated, and tracking response and privacy protection linkage decision-making based on abnormal illumination event semantics are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of video surveillance privacy protection, and more specifically, it relates to a method and system for automatically identifying and blocking sensitive areas deployed with high-position cameras. Background Art

[0002] In the high-position camera surveillance scenario, the residential windows in the surveillance area present a regular light pattern during normal periods, but abnormal light changes will occur during emergencies. When the surveillance system performs security tracking tasks, it is necessary to track the movement trajectories of suspicious targets.

[0003] The prior art establishes a baseline distribution through historical light data, detects abnormal light events deviating from the baseline in real time and classifies the patterns. At the same time, based on the conflict coordination strategy matrix of the tracking task priority and the privacy risk level of the sensitive area, the blocking adjustment plan is determined when the tracking target enters the sensitive area.

[0004] However, different types of abnormal light events have different semantic meanings. Frequent flashing patterns may indicate an emergency call for help, sudden long-on patterns may indicate private activities, and multi-window linkage patterns may indicate social scenarios such as family gatherings. Although the prior art can classify abnormal patterns and coordinate tracking conflicts, it uniformly implements the blocking enhancement strategy for all abnormal types and cannot judge whether to prioritize monitoring attention requirements or privacy protection requirements according to the abnormal semantics. The resulting technical problems are: emergency events are delayed in response due to excessive blocking, or private activities are inappropriately exposed due to tracking tasks. Summary of the Invention

[0005] The present invention provides a method and system for automatically identifying and blocking sensitive areas deployed with high-position cameras, which solves the technical problems in the related art that the surveillance system cannot intelligently adjust the privacy protection strategy according to the semantic meaning of light abnormal events, resulting in untimely response to emergency events or privacy leakage of private activities.

[0006] The present invention provides a method for automatically identifying and blocking sensitive areas deployed with high-position cameras, including: Obtaining the real-time video frame sequence collected by the high-position camera, sampling the light intensity of each window area frame by frame to generate a real-time light intensity sequence, calculating the standardized deviation score of the real-time light intensity sequence from the historical baseline distribution, and marking the window area with the standardized deviation score exceeding the preset threshold as an abnormal light event; Extracting the time-domain features of the real-time light intensity sequence of the abnormal light event, inputting the time-domain features into a time-series feature classifier, and outputting the abnormal event type label and the abnormal duration feature; Input the abnormal event type label into the event semantic inference module, perform semantic transformation based on the preset abnormal-semantic mapping rules, and output the event semantic classification result. The event semantic classification result includes potential emergency events, potential private activities, and uncertain events. Obtain the target list of active tracking tasks in the current monitoring system, calculate the spatial intersection of the predicted motion trajectory of each tracking target and the abnormal window area, and generate a set of potential conflict targets; Adjust the coordination parameters of the abnormal window area based on the event semantic classification results, increase the monitoring priority weight for potential emergency events, and reduce the authorization level for opening tracking channels for potential private activities. The coordination parameters and the tracking priority of potential conflict targets are substituted into the conflict coordination strategy matrix to calculate the coordination scheme. Based on the coordination scheme, a semantically aware occlusion mask is generated and applied to real-time video frames.

[0007] Furthermore, a multi-timescale fusion process is introduced when calculating the standardized deviation score. The standardized deviation score is calculated in both the short-term and long-term windows. The deviation scores of the short-term and long-term windows are then weighted and fused to obtain the multi-scale fused deviation score. The multi-scale fused deviation score is equal to the product of the short-term window weight coefficient and the short-term window deviation score, plus the product of the complement of the short-term window weight coefficient and the long-term window deviation score.

[0008] Furthermore, the temporal features include the frequency of light intensity change, the amplitude of light intensity change, the duration of light intensity change, the rising edge feature of light intensity, and the falling edge feature of light intensity; the abnormal event type labels include frequent flashing mode, sudden constant light mode, and multi-window linkage mode.

[0009] Furthermore, the temporal feature classifier is implemented using an LSTM neural network. The input of the LSTM neural network is a temporal feature sequence, and the output layer uses a fully connected layer and a softmax activation function to map the hidden state to the probability distribution of abnormal event type labels. When the classification confidence is lower than the preset confidence threshold, the abnormal lighting event is marked as a pending classification state, and the classification process is re-processed after the observation window is extended.

[0010] Furthermore, the anomaly-semantic mapping rules include: Frequent flashing patterns are mapped to potential emergency event semantics, sudden constant light patterns are mapped to potential private activity semantics, and multi-window linkage patterns are mapped to uncertain event semantics. When performing semantic mapping, the event semantic inference module integrates time period context information to obtain the time period identifier of the abnormal lighting event, and uses the time period identifier and the abnormal event type label together as mapping input.

[0011] Furthermore, when calculating the spatial intersection of the predicted motion trajectory of each tracked target and the abnormal window region, the length of the trajectory prediction time window is dynamically adjusted according to the motion speed of the tracked target; the trajectory prediction time window length is equal to the baseline prediction time window length multiplied by the speed adjustment factor, and the speed adjustment factor is equal to 1 plus the product of the speed adjustment coefficient and the ratio of the current motion speed of the tracked target to the reference speed.

[0012] Furthermore, the adjusted monitoring priority weight is equal to the baseline monitoring priority weight plus the weight adjustment based on event semantics plus the weight adjustment based on abnormal duration characteristics. For potential emergency events, the weight adjustment based on event semantics is positive; for potential private activities, the weight adjustment based on event semantics is negative. When the abnormal duration characteristic exceeds the duration threshold of the corresponding semantic, the weight adjustment based on the abnormal duration characteristic is positive and increases with the increase of the abnormal duration characteristic.

[0013] Furthermore, the adjusted tracking channel opening authorization level is equal to the base authorization level minus the level adjustment based on event semantics minus the level adjustment based on the abnormal duration characteristic; for potentially private activities, the level adjustment based on event semantics is positive to reduce the tracking channel opening authorization level; for potentially emergency events, the level adjustment based on event semantics is negative to increase the tracking channel opening authorization level.

[0014] Furthermore, the coordination scheme includes four strategies: maintaining occlusion, reducing occlusion intensity, opening a tracking channel, and delaying occlusion activation. The process of generating a semantically aware occlusion mask according to the coordination scheme includes: for the maintaining occlusion strategy, keeping the occlusion mask of the abnormal window region unchanged; for the reducing occlusion intensity strategy, adjusting the transparency parameter of the occlusion mask to a lower value; for the opening a tracking channel strategy, setting a transparent channel region along the predicted motion trajectory of the tracking target in the occlusion mask; and for the delaying occlusion activation strategy, setting the delay time parameter of the occlusion mask.

[0015] This invention provides an automatic sensitive area identification and occlusion system for high-position camera deployment, comprising: The anomaly detection module is used to acquire real-time video frame sequences captured by a high-position camera, sample the light intensity of each window area frame by frame to generate a real-time light intensity sequence, calculate the standardized deviation score, and mark abnormal lighting events. The feature classification module is used to extract the temporal features of abnormal lighting events and output the abnormal event type label and abnormal duration features through the temporal feature classifier. The semantic inference module is used to perform semantic transformation on abnormal event type labels based on the abnormal-semantic mapping rules and output the event semantic classification results. The conflict identification module is used to obtain a list of targets for active tracking tasks, calculate the spatial intersection of the predicted motion trajectory and the abnormal window area, and generate a set of potential conflict targets. The parameter adjustment module is used to adjust the monitoring priority weight and the authorization level for opening tracking channels based on the event semantic classification results; The occlusion generation module is used to input coordination parameters and tracking priorities into the conflict coordination strategy matrix to calculate the coordination scheme, generate semantically aware occlusion masks, and apply them to real-time video frames.

[0016] The beneficial effects of this invention are as follows: This invention transforms the type labels of abnormal lighting events into semantic classification results with decision-making guidance through an event semantic inference module. Based on these classification results, conflict coordination parameters are adjusted differentially, enabling the generation of coordination schemes to perceive the semantic meaning of abnormal lighting events. Addressing the problem of existing technologies uniformly applying occlusion enhancement strategies to all abnormal lighting events, the semantic-driven coordination parameter adjustment mechanism in this invention can distinguish the actual needs of different abnormality types. For potential emergency events corresponding to frequent flickering patterns, the coordination scheme tends to reduce occlusion intensity or open tracking channels by increasing monitoring priority. For potential private activities corresponding to sudden constant illumination patterns, the coordination scheme tends to maintain or strengthen occlusion by lowering the authorization level for opening tracking channels. This solves the technical problem of delayed response to emergency events due to excessive occlusion or improper exposure of private activities due to tracking tasks, achieving the technical effect of linking tracking response and privacy protection decisions based on the semantics of abnormal lighting events. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for automatic identification and occlusion of sensitive areas using a high-position camera, according to the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses an automatic identification and occlusion method for sensitive areas deployed by a high-position camera, such as... Figure 1 As shown, it includes the following steps: Step 1: Collect illumination data and detect abnormal illumination events.

[0020] The system acquires real-time video frame sequences captured by a high-position camera, samples the light intensity of each window area frame by frame, and generates a real-time light intensity sequence within a sliding time window. It extracts the baseline distribution parameters for the current time period from historical light data, calculates the standardized deviation score between the real-time light intensity sequence and the baseline distribution, and marks window areas with standardized deviation scores exceeding a preset threshold as abnormal light events.

[0021] Furthermore, the aforementioned frame-by-frame sampling of illumination intensity refers to extracting the pixel block corresponding to each window area in the video frame and calculating the average brightness value of all pixels within that pixel block as the illumination intensity sampling value for that frame.

[0022] Furthermore, the length of the aforementioned sliding time window is set according to the time resolution requirements of the monitoring scenario, with a typical range of 30 seconds to 5 minutes. The sliding step size of the sliding time window is usually set to an integer multiple of the video frame rate to ensure the continuity of the time sequence.

[0023] Furthermore, the calculation of the standardized deviation score involves, for each window area, using the mean and standard deviation of the light intensity corresponding to the current time period as baseline distribution parameters based on historical illumination data. The difference between the current value of the real-time illumination intensity series and the mean light intensity is divided by the standard deviation to obtain the standardized deviation score. The standardized deviation score reflects the degree of deviation of the current illumination state from the historical baseline for the same period. Specifically, Z-score standardization is used in the calculation.

[0024] Furthermore, the time span of the aforementioned historical illumination data is determined based on the periodic characteristics of the monitoring scenario. For scenarios with obvious daily cycle patterns, the time span of historical illumination data is usually set to at least two weeks to cover the differences between weekdays and weekends; for scenarios with seasonal changes, the time span of historical illumination data can be extended to several months.

[0025] Furthermore, the aforementioned preset threshold is set according to the noise level and anomaly detection sensitivity requirements of the monitoring scenario, with a typical value range of 2 to 3.

[0026] In this embodiment, to improve the robustness of anomaly detection, a multi-timescale fusion process is introduced when calculating the standardized deviation score. Specifically, the standardized deviation score is calculated within both short and long time windows, and the short-time window deviation score and the long-time window deviation score are weighted and fused to obtain the multi-scale fused deviation score. The short-time window deviation score is used to capture rapidly changing anomalous events, while the long-time window deviation score is used to capture persistent anomalous states. The multi-scale fused deviation score can simultaneously respond to both transient and gradual anomalies.

[0027] The above multi-scale fusion deviation score The calculation formula is:

[0028] in, For multi-scale fusion deviation scores, For short-term window deviation score, For long-term window deviation scores, This is the short-time window weighting coefficient, which ranges from 0 to 1.

[0029] Furthermore, the typical length of the aforementioned short window is 30 seconds to 2 minutes, used to detect rapid change events such as light flickering; the typical length of the long window is 5 minutes to 15 minutes, used to detect gradual change events such as continuous light illumination.

[0030] Furthermore, the aforementioned short-time window weighting coefficients The sensitivity settings for transient and persistent anomalies are configured based on the monitoring scenario. For scenarios requiring rapid response, a larger short-time window weighting coefficient can be set. For scenarios requiring suppression of short-term disturbances, a smaller short-term window weighting coefficient can be set. Value, short-time window weighting coefficient The typical value range is from 0.4 to 0.7.

[0031] Step 2: Extract time-domain features and classify abnormal event types.

[0032] For detected abnormal lighting events, the temporal features of the real-time sequence of light intensity are extracted, and the temporal features are input into a temporal feature classifier to analyze the temporal pattern of light changes and output abnormal event type labels and abnormal duration features.

[0033] Furthermore, the aforementioned temporal characteristics include the frequency of light intensity changes, the amplitude of light intensity changes, the duration of light intensity changes, the rising edge characteristics of light intensity, and the falling edge characteristics of light intensity.

[0034] Furthermore, the aforementioned light intensity change frequency refers to the number of times a complete cycle of light intensity jumping from a low value to a high value and then falling back to a low value occurs within a sliding time window; the light intensity change amplitude refers to the difference between the peak and valley values ​​of light intensity during a single change; the duration of light intensity change refers to the cumulative time that light intensity remains in an abnormal state; the rising edge characteristic of light intensity refers to the rate of change of light intensity from a low value to a high value, and the falling edge characteristic of light intensity refers to the rate of change of light intensity from a high value back to a low value.

[0035] Furthermore, the aforementioned abnormal event type labels include frequent flickering mode, sudden constant light mode, and multi-window linkage mode. Among them, frequent flickering mode refers to a high-frequency alternation of light intensity between light and dark within a short period of time; sudden constant light mode refers to a sudden jump in light intensity from a low value to a high value and maintaining it during an unexpected period; and multi-window linkage mode refers to the synchronous change in light status in multiple window areas of the same residence within a similar time period.

[0036] Furthermore, the criteria for determining the above-mentioned frequent flickering mode is that the frequency of light intensity changes exceeds a preset frequency threshold, with a typical threshold of more than 3 times per minute; the criteria for determining the sudden continuous light mode is that the light intensity reaches a high value state with a duration exceeding a preset duration threshold during an unusual period, with a typical duration threshold of more than 5 minutes; and the criteria for determining the multi-window linkage mode is that at least two window areas in the same residence simultaneously experience abnormal light events within a time interval less than a preset linkage time threshold, with a typical linkage time threshold of 2 minutes.

[0037] Furthermore, the aforementioned abnormal duration characteristic refers to the cumulative time from the first detection of an abnormal illumination event where the standardized deviation score exceeds a preset threshold until the standardized deviation score recovers to below the preset threshold.

[0038] Furthermore, the aforementioned temporal feature classifier is implemented using an LSTM neural network. The input to the temporal feature classifier is a temporal feature sequence, which includes the frequency, amplitude, duration, rising edge, and falling edge features of light intensity changes at each moment within the sliding time window. The output of the temporal feature classifier is the probability distribution of the abnormal event type label and the corresponding classification confidence. The LSTM neural network captures the temporal dependencies of light changes through temporal modeling. The output layer of the LSTM neural network uses a fully connected layer and a softmax activation function to map the hidden states to the probability distribution of abnormal event type labels, selecting the category with the highest probability as the abnormal event type label.

[0039] Furthermore, the aforementioned LSTM neural network structure includes an input layer, an LSTM hidden layer, and an output layer. The input layer receives the temporal feature sequence, and the LSTM hidden layer contains multiple LSTM units. Each LSTM unit controls the flow of information through a forget gate, an input gate, and an output gate, capturing long-term dependencies in the temporal feature sequence. The fully connected layer of the output layer maps the hidden state of the last time step of the LSTM hidden layer to a vector of the same dimension as the number of anomalous event type labels. The softmax activation function transforms the dimensional vector into a probability distribution.

[0040] Furthermore, the training process of the aforementioned LSTM neural network uses labeled historical anomalous lighting event data as the training set. Each sample in the training set contains a temporal feature sequence and a corresponding anomalous event type label. The training process of the LSTM neural network uses the cross-entropy loss function to measure the difference between the predicted probability distribution and the true label. The network parameters are updated through the backpropagation algorithm and the gradient descent optimizer to minimize the cross-entropy loss function value.

[0041] Furthermore, to improve the reliability of the classification results, the temporal feature classifier outputs classification confidence scores along with the abnormal event type labels. When the classification confidence score is lower than a preset confidence threshold, the abnormal lighting event is marked as pending classification, and the classification process is re-executed after extending the observation window.

[0042] Furthermore, the classification confidence score mentioned above is the maximum probability value in the probability distribution output by the softmax function. The classification confidence score reflects the degree of certainty that the temporal feature classifier is about the current classification result. The preset confidence score threshold is set according to the balance requirements between classification accuracy and response timeliness, and the typical range of the confidence score threshold is 0.6 to 0.8.

[0043] Furthermore, the aforementioned extended observation window refers to increasing the length of the sliding time window by a preset extension duration. The typical extension duration is 50% to 100% of the original sliding time window length, in order to obtain more temporal feature information, re-extract temporal features, and input them into the temporal feature classifier for classification.

[0044] Step 3: Map exception types and infer event semantics.

[0045] Input the abnormal event type label into the event semantic inference module. Based on the preset abnormal-semantic mapping rules, the abnormal event type label is semantically transformed, and the event semantic classification result is output. The event semantic classification result includes three categories: potential emergency events, potential private activities, and uncertain events.

[0046] Furthermore, the specific forms of the above-mentioned anomaly-semantic mapping rules are as follows: frequent flashing patterns are mapped to the semantics of potential emergency events, based on the fact that frequent flashing may indicate that residents are calling for help through light signals; sudden continuous illumination patterns are mapped to the semantics of potential private activities, based on the fact that continuous illumination during irregular periods may indicate that residents are engaged in private activities that they do not want to be observed; and multi-window linkage patterns are mapped to the semantics of uncertain events, based on the fact that simultaneous changes in multiple windows may correspond to various possible scenarios such as family gatherings or equipment malfunctions.

[0047] In this embodiment of the application, in order to improve the accuracy of semantic inference, the event semantic inference module integrates time period context information when performing semantic mapping.

[0048] Specifically, the time period identifier of the abnormal lighting event is obtained, and the time period identifier and the abnormal event type label are used together as the mapping input. For example, the sudden constant light pattern in the middle of the night is given a higher semantic confidence of private activity, while the frequent flashing pattern in the early morning is given a higher semantic confidence of emergency event.

[0049] Furthermore, the aforementioned time period markings are based on the time when the abnormal light event occurs. Typical time period markings include: early morning period from 6:00 to 9:00, daytime period from 9:00 to 18:00, evening period from 18:00 to 22:00, late night period from 22:00 to 2:00 the next day, and early morning period from 2:00 to 6:00.

[0050] Furthermore, the method for fusing the aforementioned time-period context information is as follows: In the anomaly-semantic mapping rule, different semantic confidence weights are assigned to the labels of anomalous event types for different time periods. For example, the semantic confidence weight for a sudden, continuous light pattern indicating private activity during the late night is set to a high value, while the semantic confidence weight for a frequent flashing pattern indicating an emergency event during the early morning is also set to a high value. The event semantic classification result output by the anomaly-semantic mapping rule selects the semantic category with the highest semantic confidence weight.

[0051] Step 4: Obtain the tracking task and identify potential conflict targets.

[0052] Obtain the target list of active tracking tasks in the current monitoring system. The target list includes the target identifier, tracking priority, real-time location coordinates and predicted motion trajectory of each tracking target. Calculate the spatial intersection between the predicted motion trajectory of each tracking target and the abnormal window area. Mark the tracking targets whose predicted motion trajectory has spatial intersection with the abnormal window area as potential conflict targets and generate a set of potential conflict targets.

[0053] Furthermore, the aforementioned target identifier is a number used to uniquely identify the tracked target in the tracking task; the tracking priority is a value set according to the threat level or importance of the tracked target, with a higher tracking priority value indicating a higher priority; the real-time position coordinates are the pixel coordinates or world coordinates of the tracked target in the current video frame; and the predicted motion trajectory is a sequence of future positions predicted based on the tracked target's historical motion trajectory.

[0054] Furthermore, the above-mentioned predicted motion trajectory is generated using a Kalman filter algorithm or a deep learning trajectory prediction model, which estimates the future position sequence within a preset time window based on the historical position coordinate sequence and motion speed of the tracked target.

[0055] Furthermore, the calculation of the aforementioned spatial intersection refers to expanding the abnormal window area to include the abnormal window and its surrounding sensitive monitoring area, and determining whether the predicted movement trajectory of the tracked target enters this spatial range within a preset future time window. If any point on the predicted movement trajectory falls within this spatial range, then spatial intersection is determined to exist.

[0056] Furthermore, the aforementioned abnormal window area is expanded as follows: based on the boundary of the abnormal window area, a preset buffer distance is extended outward to form the expanded spatial range. The preset buffer distance is set according to the spatial scale of the monitoring scene and privacy protection requirements, and the typical value range of the buffer distance is 50% to 200% of the width of the abnormal window area.

[0057] In this embodiment, to predict conflicts in advance, a trajectory prediction time window parameter is introduced when calculating spatial intersection. The length of the trajectory prediction time window is dynamically adjusted according to the movement speed of the tracked target. A longer trajectory prediction time window is used for high-speed moving targets, and a shorter trajectory prediction time window is used for low-speed moving targets, thereby controlling the false positive rate while ensuring the timeliness of the prediction.

[0058] Specifically, the length of the trajectory prediction time window The calculation formula is:

[0059] in, The length of the trajectory prediction time window. As the baseline prediction time window length, To track the target's current speed, For reference speed, This is the speed adjustment coefficient.

[0060] Furthermore, the aforementioned baseline prediction time window length The baseline prediction time window length is set based on the spatial scale of the monitoring scene and the response time requirements of the tracking task. The typical value range is 5 seconds to 30 seconds; reference speed Reference velocity is used to measure the typical motion speed of a target being tracked in a monitoring scenario. It is usually set to the median of movement speed in historical tracking data; speed adjustment factor The degree to which motion speed affects the trajectory prediction time window, and the speed adjustment coefficient. The typical value range is from 0.3 to 0.8.

[0061] Step 5: Adjust the coordination parameters according to the event semantics.

[0062] The coordination parameters of the abnormal window region in the conflict coordination process are adjusted based on the event semantic classification results. For abnormal window regions whose event semantic classification results indicate potential emergency events, the monitoring priority weight of the abnormal window region in the coordination parameters is increased; for abnormal window regions whose event semantic classification results indicate potential private activities, the authorization level for opening the tracking channel in the coordination parameters is decreased; for abnormal window regions whose event semantic classification results indicate uncertain events, the coordination parameters of the abnormal window region are maintained at the default value.

[0063] Furthermore, the aforementioned monitoring priority weight refers to the degree of priority that the monitoring system pays attention to abnormal window areas in the conflict coordination calculation. The higher the monitoring priority weight, the more the coordination result tends to reduce occlusion to facilitate monitoring. The aforementioned tracking channel opening authorization level refers to the degree of authorization allowed for the tracking task to open a tracking channel in the abnormal window area. The lower the tracking channel opening authorization level, the more the coordination result tends to maintain or strengthen occlusion to protect privacy.

[0064] In this embodiment, to achieve fine-grained adjustment of coordination parameters, the abnormal duration characteristic is incorporated into the coordination parameter adjustment process. Specifically, for potential emergency event semantics, when the abnormal duration characteristic exceeds a preset emergency duration threshold, the monitoring priority weight is further increased; for potential private activity semantics, when the abnormal duration characteristic exceeds a preset private duration threshold, the authorization level for opening tracking channels is further reduced.

[0065] Specifically, the adjusted monitoring priority weights The calculation formula is:

[0066] in, The adjusted monitoring priority weights, Based on the benchmark monitoring priority weight, This is a weight adjustment based on event semantics. This is the weight adjustment amount based on the abnormal duration feature.

[0067] Furthermore, regarding potential emergencies, Take a positive value; for potentially private activities, Take negative values; for uncertain events, Take zero.

[0068] Furthermore, the above The typical range of values ​​is: potential emergency events The value is the baseline monitoring priority weight. 20% to 50% of potential private activities The value is the baseline monitoring priority weight. -20% to -50%.

[0069] Furthermore, The calculation method is as follows: when the duration of the anomaly exceeds the duration threshold of the corresponding semantic meaning, Take a positive value and increase it as the duration of the anomaly increases; otherwise... Take zero.

[0070] Furthermore, the typical values ​​for the emergency duration thresholds mentioned above are 2 to 5 minutes, and the typical values ​​for the private duration thresholds are 5 to 10 minutes.

[0071] Similarly, the adjusted tracking channels open up authorization levels. The calculation formula is:

[0072] in, Authorization levels will be established for the adjusted tracking channels. As the baseline authorization level, This is a level adjustment based on event semantics. This is the level adjustment amount based on the duration of the anomaly.

[0073] Furthermore, regarding potentially private activities, Take a positive value to lower the authorization level for opening a tracking channel; for potential emergencies, Taking negative values ​​increases the authorization level for opening tracking channels; for uncertain events, Take zero.

[0074] Furthermore, the above The typical range of values ​​is: potential private activities Values ​​range from 1 to 3 levels; potential emergency events The value ranges from -1 to -3 levels. The authorization level for tracing channel opening is usually represented by a discrete integer; the higher the value, the higher the level of authorization.

[0075] Furthermore, The calculation method is as follows: when the duration of the anomaly exceeds the duration threshold of the corresponding semantic meaning, Take a positive value and increase it as the duration of the anomaly increases; otherwise... Take zero.

[0076] Step 6: Calculate the coordination scheme and generate the occlusion mask.

[0077] The semantically adjusted coordination parameters and the tracking priorities of each tracking target in the potential conflict target set are substituted into the conflict coordination strategy matrix. A coordination scheme is calculated for each conflict item formed by the tracking target and the abnormal window region. The coordination scheme includes four strategies: maintaining occlusion, reducing occlusion intensity, opening a tracking channel, and delaying the occlusion effect. A semantically aware occlusion mask is generated according to the coordination scheme. The semantically aware occlusion mask is applied to real-time video frames to output a privacy-preserving video stream with event semantic linkage.

[0078] Furthermore, the aforementioned conflict coordination strategy matrix is ​​a decision mapping structure with monitoring priority weight and tracking priority as input dimensions and coordination scheme as output. When the monitoring priority weight is higher than the tracking priority, the coordination scheme tends to reduce the occlusion intensity or open a tracking channel; when the authorization level for opening a tracking channel is lower than the tracking priority, the coordination scheme tends to maintain occlusion or delay the occlusion from taking effect.

[0079] Furthermore, the decision-making logic of the conflict coordination strategy matrix is ​​as follows: Calculate the difference between monitoring priority weight and tracking priority, and the difference between tracking channel opening authorization level and tracking priority; when monitoring priority weight is significantly higher than tracking priority and tracking channel opening authorization level is not lower than tracking priority, select the tracking channel opening strategy; when monitoring priority weight is higher than tracking priority but tracking channel opening authorization level is lower than tracking priority, select the strategy of reducing occlusion intensity; when monitoring priority weight is not higher than tracking priority and tracking channel opening authorization level is significantly lower than tracking priority, select the strategy of maintaining occlusion; when monitoring priority weight is not higher than tracking priority and tracking channel opening authorization level is slightly lower than tracking priority, select the strategy of delaying occlusion activation.

[0080] Furthermore, the process of generating semantically aware occlusion masks based on the coordination scheme includes: converting the strategy types in the coordination scheme into corresponding occlusion parameter configurations. For the occlusion maintenance strategy, the occlusion mask for the abnormal window region remains unchanged; for the occlusion intensity reduction strategy, the transparency parameter of the occlusion mask is adjusted to a lower value; for the tracking channel opening strategy, a transparent channel region is set in the occlusion mask along the predicted motion trajectory of the tracking target; for the delayed occlusion activation strategy, the delay time parameter of the occlusion mask is set. The converted occlusion parameter configuration is applied to the corresponding region of the real-time video frame to generate a semantically aware occlusion mask.

[0081] Furthermore, the transparency parameter of the aforementioned occlusion mask ranges from 0 to 1. A transparency parameter of 0 indicates complete transparency (no occlusion), while a transparency parameter of 1 indicates complete opacity (complete occlusion). The occlusion maintenance strategy keeps the original transparency parameter unchanged; the occlusion reduction strategy reduces the transparency parameter to a preset reduction level, typically 20% to 50% of the original transparency parameter.

[0082] Furthermore, the above-mentioned transparent channel area is set as follows: based on the predicted motion trajectory of the tracking target, a transparent area with a width of a preset channel width is set in the occlusion mask along the path of the predicted motion trajectory. The preset channel width is set according to the size of the tracking target and the tracking accuracy requirements. The typical value of the preset channel width is 1.2 to 2 times the width of the tracking target bounding box.

[0083] Furthermore, the aforementioned delay time parameter is the waiting time before the occlusion mask takes effect. The delayed occlusion strategy sets the delay time parameter to a preset delay duration. The typical delay duration is 2 to 10 seconds, which is used to provide a buffer time when the tracking target is about to enter the sensitive area but has not yet entered.

[0084] Furthermore, the process of applying the occlusion parameter configuration to the real-time video frame is as follows: based on the transparency parameter and spatial location information of the occlusion mask, the pixels in the corresponding area of ​​the real-time video frame are subjected to transparency blending processing. Specifically, the original pixel values ​​and the occlusion color are weighted and blended according to the transparency parameter to generate the occluded pixel values. For the transparent channel area, the original pixel values ​​are kept unchanged.

[0085] In this embodiment, to support post-event auditing and strategy optimization, while generating the semantically aware occlusion mask, the event semantic classification result, coordination parameter adjustment record, coordination scheme, and corresponding tracking target information are written to the audit log. The records in the audit log include timestamps, abnormal window area identifiers, event semantic classification results, coordination scheme types, and coordination scheme execution status.

[0086] Furthermore, the audit logs mentioned above are stored in a structured data format. Each record in the audit log contains the following fields: timestamp (recording the precise time when the abnormal lighting event occurred), abnormal window area identifier (recording the window area number where the abnormal lighting event occurred), event semantic classification result (recording the inferred semantic category), coordination scheme type (recording the selected strategy type), coordination scheme execution status (recording whether the coordination scheme was successfully applied), and tracking target information (recording the tracking target identifier and tracking priority).

[0087] Furthermore, audit logs are used for post-event analysis to assess the rationality of coordination decisions. By statistically analyzing the distribution of coordination schemes under different event semantic categories, the effectiveness of semantic-driven coordination mechanisms is evaluated, and anomaly-semantic mapping rules and coordination parameter adjustment strategies are optimized based on actual response effects.

[0088] This implementation uses an event semantic inference module to convert the type labels of abnormal lighting events into event semantic classification results with decision-making guidance, and adjusts the conflict coordination parameters differently based on the event semantic classification results, so that the generation of coordination schemes can perceive the semantic meaning of abnormal lighting events.

[0089] To address the issue of existing technologies uniformly applying occlusion enhancement strategies to all abnormal lighting events, the semantic-driven coordination parameter adjustment mechanism in this implementation can differentiate the actual needs of different anomaly types. For potential emergency events corresponding to frequent flickering patterns, the coordination scheme is tilted towards reducing occlusion intensity or opening tracking channels by increasing the monitoring priority weight, thereby ensuring the monitoring system's ability to attend to potential distress signals. For potential private activities corresponding to sudden constant illumination patterns, the coordination scheme is tilted towards maintaining or strengthening occlusion by lowering the authorization level for opening tracking channels, thereby ensuring the privacy protection of residents' private activities.

[0090] Therefore, this implementation overcomes the shortcomings of anomaly detection and conflict coordination running independently, which cannot handle differences in semantic-level decision-making requirements. It realizes the linkage decision-making between tracking response and privacy protection based on the semantics of abnormal lighting events, avoiding the problem of delayed response to emergency events due to excessive occlusion or improper exposure of private activities due to tracking tasks.

[0091] A large residential community has deployed a high-position camera surveillance system responsible for security monitoring of the community's public areas. At 22:35 on January 21, 2026, the surveillance system was tracking a suspicious person loitering within the community (target ID: TARGET_20260121_089), who was marked as a medium-priority tracking task (priority level 5). Simultaneously, an abnormal light change occurred in the bedroom window of apartment 502, unit 2, building 3 (window ID: WIN_03_02_502_BR): the bedroom, which was previously dark, suddenly turned on at 22:36 and remained bright for the next 3 minutes. The surveillance system needs to determine the semantic meaning of this abnormal light event and, as the predicted trajectory of the suspicious person is about to pass near the window of apartment 502, decide whether to maintain privacy obstruction of the window area or open an observation path for the tracking task.

[0092] Example of implementing core step 1: Acquire illumination data and detect abnormal illumination events.

[0093] The monitoring system samples the illumination intensity of the WIN_03_02_502_BR window area frame by frame and performs multi-scale fusion analysis using a 60-second short window and a 10-minute long window.

[0094] Table 1 Input data for step 1:

[0095] Table 2 Output data for step 1:

[0096] Calculation process: The short-term window deviation score is obtained by dividing the difference between the current illuminance (142.8) and the historical mean (15.3) by the standard deviation (8.7). The long-term window deviation score is obtained through smoothing calculations over a longer time range. The multi-scale fusion deviation score is calculated as follows:

[0097] because (Preset threshold) The system determines that an abnormal lighting event has occurred in the WIN_03_02_502_BR window area.

[0098] Example of implementing core step 2: extracting time-domain features and classifying abnormal event types.

[0099] The system extracts temporal features within a 3-minute observation window from detected abnormal lighting events and inputs them into an LSTM temporal feature classifier for pattern recognition.

[0100] Table 3 Input data for step 2:

[0101] Table 4 Output data for step 2:

[0102] After analyzing the temporal features, the LSTM classifier outputs a probability distribution. Since the light intensity change frequency is only 0.33 times / minute (far below the frequent flicker threshold of 3 times / minute) and the duration reaches 180 seconds (exceeding half of the sudden long-on threshold of 5 minutes but still under continuous observation), and only a single window changes (not meeting the multi-window linkage condition), the sudden long-on mode obtains the highest probability of 0.85. The classification confidence score of 0.85 exceeds the preset threshold of 0.6, indicating the classification result is valid.

[0103] Implementation example of core step 3: Mapping exception types and inferring event semantics.

[0104] The system combines the sudden continuous light pattern label with the context information of the late night period into the event semantic inference module for semantic conversion.

[0105] Table 5 shows the input and output data for step 3:

[0106] According to the anomaly-semantic mapping rule, the basic semantics of the sudden constant lighting pattern is potential private activity. Combined with the context of late-night time, the system increases the semantic confidence weight of private activity to 1.35 (higher than the baseline value of 1.0), and finally confirms that the event semantics are potential private activity, with a semantic inference confidence of 0.91.

[0107] Example of implementing core step 4: Obtain the tracking task and identify potential conflict targets.

[0108] The system obtains information about the currently active tracking task TARGET_20260121_089 and calculates the spatial intersection of its predicted trajectory and the abnormal window region.

[0109] Table 6 shows the input and output data for step 4:

[0110] The calculation process for the trajectory prediction time window length:

[0111] Based on the target's current position (38.2, 125.7) and its speed of 1.2 m / s, it is predicted that the target will reach position (41.8, 127.9) in 26.25 seconds. The distance between this position and the center of the anomaly window area (42.5, 128.3) is:

[0112] because (Extended range radius) It was determined that there was spatial intersection, and TARGET_20260121_089 was marked as a potential conflict target.

[0113] Implementation example of core step 5: Adjusting coordination parameters based on event semantics.

[0114] The system adjusts the coordination parameters of the WIN_03_02_502_BR window area based on the semantics of potential private activities and the abnormal duration of 180 seconds.

[0115] Table 7 shows the input and output data for step 5:

[0116] Adjusted monitoring priority weight calculation process:

[0117] Because the event's semantics indicate a potentially private activity. Take a negative value of -1.2 (40% of the baseline value of 3.0). Since the abnormal duration of 180 seconds does not exceed the privacy duration threshold of 300 seconds, .

[0118] Adjusted tracking channel opening authorization level calculation process:

[0119] Because the event's semantics indicate a potentially private activity. A positive value of 2 is used to lower the authorization level. Since the duration did not exceed the threshold, .

[0120] Implementation example of core step 6: Calculate the coordination scheme and generate the occlusion mask.

[0121] The system substitutes the adjusted coordination parameters and the tracking target priority into the conflict coordination strategy matrix, calculates the coordination scheme, and generates a semantically aware occlusion mask.

[0122] Input data for step 6 in Table 8:

[0123] Table 9 Output data for step 6:

[0124] Coordination Decision Logic: Monitoring Priority Weighting The tracking priority is below level 5 (normalized to 5.0), and the tracking channel opening authorization level is [not specified]. The level is significantly lower than the tracking priority level by 5, with a difference of 3 levels. According to the conflict coordination strategy matrix, when the monitoring priority weight is not higher than the tracking priority and the authorization level is significantly lower than the tracking priority, the maintenance occlusion strategy is selected.

[0125] The generated occlusion parameters are configured as follows: maintain the occlusion transparency parameter of the WIN_03_02_502_BR window area at 0.88 (close to complete occlusion), do not set a transparency channel area, and set the delay time parameter to 0 seconds (effective immediately). The system applies this occlusion mask to real-time video frames to ensure privacy protection for the windows of bedroom 502 as the tracking target passes by.

[0126] The data flow throughout the implementation process reflects the complete link from raw illumination sampling to final occlusion decision: Illumination sampling and anomaly detection: The system extracts the illumination intensity of 142.8 in the WIN_03_02_502_BR window area from the video frame, calculates the multi-scale fusion deviation score of 13.96 with the historical baseline parameters (mean 15.3, standard deviation 8.7), and triggers anomaly detection.

[0127] Temporal feature extraction and classification: After the anomaly is triggered, the system extracts temporal features (change frequency of 0.33 times / minute, duration of 180 seconds, etc.) within a 180-second observation window, and the LSTM classifier outputs a sudden continuous light pattern label (confidence level of 0.85).

[0128] Semantic inference and context fusion: The sudden continuous light pattern label is combined with the context of the late night period and converted into the semantics of potential private activities through the anomaly-semantic mapping rule (confidence 0.91).

[0129] Tracking task conflict identification: The system obtains the motion parameters (speed 1.2 m / s) of the tracking target TARGET_20260121_089, calculates the trajectory within the 26.25-second prediction time window, and finds that the distance between the predicted position (41.8, 127.9) and the window area (42.5, 128.3) is 0.81 meters, and determines that there is a spatial intersection.

[0130] Coordination parameter semantic adjustment: Based on the semantics of potential private activities, the system reduces the monitoring priority weight from 3.0 to 1.8 and the authorization level from level 4 to level 2, reflecting the response to privacy protection needs.

[0131] Coordination Decision and Occlusion Generation: The comparison between the adjusted coordination parameters (monitoring priority weight 1.8, authorization level 2) and tracking priority level 5 drives the conflict coordination strategy matrix to select the occlusion scheme, and finally generates an occlusion mask with an transparency of 0.88, which is applied to the video stream.

[0132] The key to data flow lies in the fact that the 180-second anomaly duration output in step 2 is used by both the semantic inference in step 3 and the coordination parameter adjustment in step 5. The potential conflict target set output in step 4 and the adjusted coordination parameters output in step 5 are input into the decision matrix in step 6, ensuring the integrity and consistency of semantic information transmitted from the anomaly detection end to the occlusion control end.

[0133] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for automatic identification and occlusion of sensitive areas in high-position camera deployment, characterized in that, Includes the following steps: The system acquires real-time video frame sequences captured by a high-position camera, samples the light intensity of each window area frame by frame to generate a real-time light intensity sequence, calculates the standardized deviation score between the real-time light intensity sequence and the historical baseline distribution, and marks window areas with standardized deviation scores exceeding a preset threshold as abnormal light events. Extract the temporal features of the real-time sequence of light intensity of abnormal lighting events, input the temporal features into a time-series feature classifier, and output the abnormal event type label and abnormal duration features; Input the abnormal event type label into the event semantic inference module, perform semantic transformation based on the preset abnormal-semantic mapping rules, and output the event semantic classification result. The event semantic classification result includes potential emergency events, potential private activities, and uncertain events. Obtain the target list of active tracking tasks in the current monitoring system, calculate the spatial intersection of the predicted motion trajectory of each tracking target and the abnormal window area, and generate a set of potential conflict targets; Adjust the coordination parameters of the abnormal window area based on the event semantic classification results, increase the monitoring priority weight for potential emergency events, and reduce the authorization level for opening tracking channels for potential private activities. The coordination parameters and the tracking priority of potential conflict targets are substituted into the conflict coordination strategy matrix to calculate the coordination scheme. Based on the coordination scheme, a semantically aware occlusion mask is generated and applied to real-time video frames.

2. The method for automatic identification and occlusion of sensitive areas using a high-position camera as described in claim 1, characterized in that, Multi-timescale fusion processing is introduced when calculating the standardized deviation score. The standardized deviation score is calculated in both short-time and long-time windows. The deviation scores in the short-time window and the long-time window are then weighted and fused to obtain the multi-scale fused deviation score. The multi-scale fusion deviation score is equal to the product of the short-term window weight coefficient and the short-term window deviation score, plus the product of the complement of the short-term window weight coefficient and the long-term window deviation score.

3. The method for automatic identification and occlusion of sensitive areas using a high-position camera as described in claim 1, characterized in that, The time-domain features include the frequency of light intensity change, the amplitude of light intensity change, the duration of light intensity change, the rising edge features of light intensity, and the falling edge features of light intensity. The abnormal event type labels include frequent flashing mode, sudden constant light mode, and multi-window linkage mode.

4. The method for automatic identification and occlusion of sensitive areas using a high-position camera as described in claim 3, characterized in that, The temporal feature classifier is implemented using an LSTM neural network. The input of the LSTM neural network is a temporal feature sequence, and the output layer uses a fully connected layer and a softmax activation function to map the hidden state to the probability distribution of abnormal event type labels. When the classification confidence is lower than the preset confidence threshold, the abnormal lighting event is marked as a pending classification state, and the classification process is re-processed after the observation window is extended.

5. The method for automatic identification and occlusion of sensitive areas using a high-position camera as described in claim 1, characterized in that, The anomaly-semantic mapping rules include: Frequent flashing patterns are mapped to potential emergency event semantics, sudden constant light patterns are mapped to potential private activity semantics, and multi-window linkage patterns are mapped to uncertain event semantics. When performing semantic mapping, the event semantic inference module integrates time period context information to obtain the time period identifier of the abnormal lighting event, and uses the time period identifier and the abnormal event type label together as mapping input.

6. The method for automatic identification and occlusion of sensitive areas using a high-position camera as described in claim 1, characterized in that, When calculating the spatial intersection of the predicted motion trajectory of each tracked target and the abnormal window region, the length of the trajectory prediction time window is dynamically adjusted according to the movement speed of the tracked target. The trajectory prediction time window length is equal to the baseline prediction time window length multiplied by the speed adjustment factor. The speed adjustment factor is equal to 1 plus the product of the speed adjustment coefficient and the ratio of the current movement speed of the tracked target to the reference speed.

7. The method for automatic identification and occlusion of sensitive areas using a high-position camera as described in claim 1, characterized in that, The adjusted monitoring priority weight equals the baseline monitoring priority weight plus the weight adjustment based on event semantics plus the weight adjustment based on the anomaly duration feature; For potential emergency events, the weight adjustment based on event semantics is positive; for potential private activities, the weight adjustment based on event semantics is negative; when the abnormal duration feature exceeds the duration threshold of the corresponding semantics, the weight adjustment based on the abnormal duration feature is positive and increases with the increase of the abnormal duration feature.

8. The method for automatic identification and occlusion of sensitive areas using a high-position camera as described in claim 1, characterized in that, The adjusted authorization level for opening a tracking channel is equal to the base authorization level minus the level adjustment based on event semantics and then minus the level adjustment based on the duration of the anomaly. For potentially private activities, the level adjustment based on event semantics is positive to lower the authorization level for opening a tracking channel; for potentially urgent events, the level adjustment based on event semantics is negative to raise the authorization level for opening a tracking channel.

9. The method for automatic identification and occlusion of sensitive areas in high-position camera deployment according to claim 1, characterized in that, The coordination scheme includes four strategies: maintaining occlusion, reducing occlusion intensity, opening tracking channels, and delaying the occlusion effect. The process of generating semantically aware occlusion masks according to the coordination scheme includes: for the occlusion maintenance strategy, keeping the occlusion mask of the abnormal window area unchanged; For the strategy of reducing occlusion intensity, the transparency parameter of the occlusion mask is adjusted to a lower value; for the strategy of opening a tracking channel, a transparent channel area is set in the occlusion mask along the predicted motion trajectory of the tracking target; for the strategy of delaying the occlusion effect, the delay time parameter of the occlusion mask is set.

10. A system for automatic identification and occlusion of sensitive areas using a high-position camera deployment, used to execute the method for automatic identification and occlusion of sensitive areas using a high-position camera deployment as described in any one of claims 1 to 9, characterized in that, include: The anomaly detection module is used to acquire real-time video frame sequences captured by a high-position camera, sample the light intensity of each window area frame by frame to generate a real-time light intensity sequence, calculate the standardized deviation score, and mark abnormal lighting events. The feature classification module is used to extract the temporal features of abnormal lighting events and output the abnormal event type label and abnormal duration features through the temporal feature classifier. The semantic inference module is used to perform semantic transformation on abnormal event type labels based on the abnormal-semantic mapping rules and output the event semantic classification results. The conflict identification module is used to obtain a list of targets for active tracking tasks, calculate the spatial intersection of the predicted motion trajectory and the abnormal window area, and generate a set of potential conflict targets. The parameter adjustment module is used to adjust the monitoring priority weight and the authorization level for opening tracking channels based on the event semantic classification results; The occlusion generation module is used to input coordination parameters and tracking priorities into the conflict coordination strategy matrix to calculate the coordination scheme, generate semantically aware occlusion masks, and apply them to real-time video frames.

Images