Security robot dog alarm response method for multi-modal perception

Abstract

The application discloses a security robot dog alarm response method for multi-modal perception, and particularly relates to the technical field of security inspection and multi-robot cooperative control, obtains visible light blocking rate, infrared temperature difference, acoustic echo time delay, gas concentration change rate and pose information of each robot dog, and forms an alarm observation sequence; determines a modal inversion boundary and an alarm confidence vector on both sides of the boundary according to the air flow direction of the roller shutter door gap; establishes an alarm shadow map in combination with the passable path, and selects a boundary node with the largest information missing degree to perform detection and crossing; corrects the alarm shadow map by using the observation difference before and after the crossing, and obtains a target alarm area; divides detection, containment and communication relay roles according to the channel occupation cost, and generates a staggered response path; the application can reduce alarm misjudgment and narrow channel congestion in a complex shielding environment.

Description

Technical Field

[0001] This invention relates to the field of security patrol and multi-robot collaborative control technology, specifically to a collaborative response method for security robot dogs based on multimodal perception. Background Technology

[0002] Current security robot dogs typically rely on visible light, infrared, and acoustic information for alarm location. However, within the fire compartments of underground parking garages, when roller shutters are partially lowered and sprinkler systems and smoke exhaust fans are activated simultaneously, water mist obscuring the view, metal reflections, and airflow through the door gaps can cause the same alarm source to exhibit temperature peak shifts, sound source echo reversals, and gas concentration lags on both sides of the roller shutter. If multiple robot dogs approach the target area directly based on a single confidence level, they are prone to concentrating on the wrong side of the roller shutter, causing passageway congestion and delaying the response to trapped personnel or battery thermal runaway points.

[0003] Therefore, how to identify modal inversion regions caused by local obstacles and respond accordingly has become an urgent technical problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a collaborative response method for security robot dogs based on multimodal perception, in order to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a collaborative response method for alarm situations using a security robot dog oriented towards multimodal perception, comprising: The visible light occlusion rate, infrared temperature difference, acoustic echo delay, gas concentration change rate and pose information collected by each robot dog are obtained to form an alarm observation sequence. Based on the alarm observation sequence and the airflow direction of the roller shutter door gap, the modal inversion boundary where the response direction of different sensing modes reverses is determined, and the alarm confidence vectors on both sides of the boundary are obtained. An alarm shadow map is established based on the alarm confidence vector and the passable path. The alarm shadow map includes shadow nodes that record risk values ​​and information missingness, and occupancy gates that record the cost of channel occupancy. Based on the aforementioned alarm shadow map, the boundary node with the largest information missing degree is selected, and at least one robot dog is controlled to perform detection and crossing. The alarm shadow map is then corrected using the observation difference before and after the crossing to obtain the target alarm area. Based on the revised alarm shadow map, each robot dog is divided into detection, control and communication relay roles, and staggered response paths are generated according to the channel occupancy cost.

[0006] Preferably, an alarm observation sequence is formed, including: Control the robot dog to approach the roller shutter door and read the pressure at the end of the foreleg, the longitudinal micro-vibration of the body and the corresponding angular velocity of the foreleg joint; The moment when the foot pressure changes to a stable bearing state, the fuselage experiences a short-term longitudinal micro-vibration, and the corresponding forefoot joint angular velocity changes from a swinging state to a convergent state is determined as the synchronization marker. The time interval between adjacent synchronization markers is defined as the landing cycle window. Various types of sensing data are associated with a unified timestamp. The sensing data and pose information within the same landing cycle window are arranged as observation records and formed into an alarm observation sequence according to the order of movement.

[0007] Preferably, a narrow strip observation area is set along the edge contour of the roller shutter door, and contour units that only experience contour breakage or water mist coverage during the current landing cycle are marked as intermittent blanking units to obtain the visible light occlusion rate; temperature sampling bands are set in the adjacent areas on the windward and leeward sides of the door gap to obtain the infrared temperature difference; coded acoustic short pulses are emitted between adjacent synchronization marks, and the acoustic echo delay is obtained based on the primary reflection from the door gap edge and the secondary reflection from the ground; a lateral sweep is performed intersecting the airflow direction of the door gap, and the gas concentration change rate is obtained based on the gas sensor recovery segment.

[0008] Preferably, determining the modal inversion boundary includes: The airflow direction at the roller shutter door gap is corrected based on the drift direction of lightweight particulate matter and the recovery section of the gas sensor. Based on the pose information, a fold sampling zone is established in the environmental map. The fold sampling zone includes a windward side branch, a leeward side branch, and a branch adjacent to the door gap connecting the two. Pair the corresponding observation records in the windward and leeward branches; The stable switching position of the arrival order of primary reflection at the edge of the door gap and secondary reflection from the ground is determined as the gap-penetrating reference interval, and continuous observation windows in which the response directions of at least two sensing modes are reversed are marked as candidate inverted segments.

[0009] Preferably, the alarm confidence vectors obtained on both sides of the boundary include: Select a verification robot dog located within the safe passage area and control the verification robot dog to perform lateral reciprocating motion along the verification path that intersects with the airflow direction of the roller shutter door gap; According to the synchronization mark, the outward observation record and the corresponding spatial position observation record in the return are interleaved and spliced, and the candidate inverted segments that only appear in one direction are eliminated. The modal inversion boundary is extracted from the candidate inverted segments that maintain the reverse of the response direction in both the round trip. The observation records on both sides of the modal inversion boundary are collected separately, and the response direction, duration and confidence level are written in a fixed order of visible light, infrared, acoustic and gas sensing to form an alarm confidence vector.

[0010] Preferably, a police incident shadow map is created, including: Project the alarm confidence vectors on both sides of the modal inversion boundary onto the passable path, and determine the direction corresponding to the perception item with high or medium confidence level as the initial alarm direction; Perform reverse backtracking along the passable path in the opposite direction of each initial alarm; When the reverse tracing path encounters a corner, roller shutter door rail, vehicle side, vehicle bottom, or fire-fighting facility base, and an observation interruption or directional shift occurs, insert a shaded node; Shadow nodes located at the same shading structure entrance or the same path turning point are merged into composite shadow nodes, and shadow nodes directly connected to the modal inverted boundary are marked as boundary nodes.

[0011] Preferably, the creation of a police incident shadow map also includes: Determine the upstream and downstream shadow nodes of each shadow node according to the reverse backtracking direction, and record the perception items that cannot be further verified by the corresponding perception mode after the path is turned as downstream missing items. Based on the pose information corresponding to the synchronization marker, the landing cycle of the missing item that is continuously crossed is converted into a continuous unchecked segment, and the information missing degree is written. Risk values ​​are written based on the retention status of upstream high-confidence sensing items before and after path reversal and the consistency of cross-modal alert directions; Path segments that cannot accommodate two robot dogs passing in parallel or safely meeting are marked as occupied gates, and the channel occupancy cost is written according to the occupancy status, the expected entry and exit order, and whether the evacuation path is cut off after the blockage.

[0012] Preferably, performing a probe crossing includes: Read the shadow nodes downstream along the shadow branches corresponding to each boundary node, and write the highest information missing level among the reachable downstream shadow nodes into the corresponding boundary node. When the information missing degree of multiple boundary nodes is the same, the detection crossing entrance is determined sequentially based on the number of landing cycles of consecutive crossing of the missing item, the number of downstream high-risk shadow nodes, and the channel occupation cost of the occupied gate. Control at least one detection robot dog to enter the shadow branch from the detection passage entrance, and set stopping points before the occlusion turning point, when passing through the occlusion turning point, and after leaving the occlusion turning point; Lateral sweeps and multimodal sampling were performed at each stop point.

[0013] Preferably, the correction of the police shading map and the acquisition of the target police area include: Using the occlusion and reversal position as a reference, the observation records downstream, in the middle and upstream are read in sequence to form a reverse staggered observation record sequence; Extraction of intermittent blanking unit position retraction, infrared temperature difference direction switching, acoustic reflection arrival sequence reset and gas recovery segment extension; When at least two observation differences appear within the same correction window covering no more than two adjacent landing cycles, a correction node is inserted, and valid observations that repeat downstream of the correction node are updated to the verified status. Update the information missingness based on the remaining missing items, update the risk value based on the consistency of cross-modal alerts, and classify the shadow nodes that continuously maintain a high risk value into the target alert area.

[0014] Preferably, generating off-peak response paths includes: Extract high-risk shadow nodes, boundary nodes, and occupied gates that are connected to the target alarm area from the revised alarm shadow map, and write the exit synchronization flag of each occupied gate in sequence along the evacuation direction; Robot dogs that can enter the target alarm area and have enough remaining power to meet the evacuation requirements are classified as detection roles; robot dogs that can reach the boundary node and do not occupy the evacuation path are classified as control roles; and robot dogs that can maintain two-way communication with both the collaborative control terminal and the detection role are classified as communication relay roles. Reverse time slot reservations are executed in descending order of channel occupancy cost; When the gate is delayed in being released, the reserved time slots that have not yet been executed are extended, and the staggered response path is updated according to the revised alarm shadow map.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention does not directly locate alarms based on single heat source peaks or gas concentration peaks. Instead, it uses the vibration of the structure where the foot lands as a multimodal synchronous marker. It establishes a gap-penetration reference by combining the arrival sequence switching of primary reflection from the door gap and secondary reflection from the ground, and identifies the reversal of the response directions of visible light, infrared, and gas sensing on both sides of the roller shutter door. Therefore, even when water spray, metal reflection, and exhaust airflow are present simultaneously, it can reduce misjudgments caused by sampling misalignment and apparent direction shifts, accurately distinguishing between genuine alarm sources and false alarm areas formed by door gap disturbances.

[0016] This invention traces the perception results from both sides of the modal inversion boundary back along the passable path to the guide rail angle, wall corner, and vehicle bottom, forming an alarm shadow map that records risk values, information gaps, and channel occupancy costs. It then controls the robot dog to cross obstructed turning points for three-point detection and passage. By recovering the verified missing items through the observation difference before and after the passage, and reserving reverse time slots based on the occupancy threshold, it avoids multiple robot dogs simultaneously entering the wrong side or blocking narrow passages, while maintaining continuous coordination between detection, control, and communication relay. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of the security robot dog alarm collaborative response method for multimodal perception according to the present invention.

[0019] Figure 2 This is a flowchart of the method for determining the modal inversion boundary according to the present invention.

[0020] Figure 3 This is a flowchart of the method for generating staggered response paths according to the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1, please refer to Figure 1 As shown in this embodiment, the collaborative response method for security robot dogs oriented towards multimodal perception includes: The system acquires information on visible light occlusion rate, infrared temperature difference, acoustic echo delay, gas concentration change rate, and pose from each robot dog, forming an alarm observation sequence.

[0023] It should be noted that this embodiment is applied to the inspection scenario of fire compartments in underground parking garages. The roller shutter door is in a half-lowered state, the sprinkler system and smoke exhaust fan have been activated, and a gap is formed between the lower edge of the roller shutter door and the ground, allowing airflow to pass through. Due to the influence of sprinkler water mist, smoke exhaust negative pressure, metal door reflection, and ground water, the area near the gap is prone to intermittent disappearance of visible light outlines, infrared heat shift, alternating acoustic reflection paths, and short-term fluctuations in gas concentration.

[0024] Each robot dog participating in the coordinated response is equipped with a visible light camera module, an infrared thermal imaging module, a speaker, a microphone array, a gas sensor, an inertial measurement unit (IMU), a joint encoder, and a foot pressure sensor. The gas sensor is used to detect at least one of carbon monoxide, particulate matter, or volatile gases. Each sensor is connected to the robot dog's local controller, which writes a unified timestamp to each sensor. The robot dog continuously updates its pose information based on the IMU, joint encoder, and a pre-built environmental map.

[0025] The robot dog approaches the roller shutter door with a low-speed gait. The controller reads the pressure changes from the pressure sensors at the ends of the forelegs and simultaneously reads the longitudinal micro-vibrations of the body detected by the inertial measurement unit and the joint angular velocities output by the corresponding foreleg joint encoders. When the foot pressure changes from a low value to a stable bearing state, the body experiences a short-term longitudinal micro-vibration, and the corresponding foreleg joint angular velocity changes from a swinging state to a convergent state, this moment is designated as the synchronization marker for the foreleg landing. If only the body experiences micro-vibration without a change in the bearing state of the feet, the controller identifies it as a pseudo-vibration caused by the impact of sprayed water droplets or splashing water on the ground and does not record it in the synchronization marker.

[0026] The controller defines the time interval between two adjacent synchronization markers as a landing cycle window, and associates the visible light image, infrared image, acoustic data, gas sensor data, and pose information within the landing cycle window with the same synchronization marker according to a unified timestamp. For pose information that does not fall exactly at the synchronization marker, the controller interpolates based on the continuous poses before and after the synchronization marker, so that each synchronization marker corresponds to a specific spatial position and fuselage orientation.

[0027] When obtaining the visible light occlusion rate, the controller first extracts the lower edge of the roller shutter door and the vertical fold line of the door body from multiple frames of visible light images, and uses both to determine the edge contour of the roller shutter door. A narrow strip observation area is set along the edge contour of the roller shutter door, and then the narrow strip observation area is divided into continuously arranged contour units. For any contour unit, if the contour unit can be identified in the previous and subsequent landing cycles, but shows brightness gradient attenuation, contour breakage, or is covered by water mist in the current landing cycle, it is marked as an intermittent blanking unit. For contour units that are missing for multiple consecutive landing cycles, they are marked as fixed missing units to eliminate the influence of door corrosion, fixed stains, or structural dents.

[0028] The controller calculates the proportion of intermittent blanking cells to effective contour cells within the current landing cycle to obtain the visible light occlusion rate. The calculation expression is as follows: In the formula, The number of effective contour units, This represents the number of intermittent blanking units. The visible light occlusion rate is defined as the visible light occlusion rate. The effective contour unit is the remaining contour unit after removing fixed missing units. Therefore, the visible light occlusion rate mainly reflects the dynamic occlusion caused by water mist in the door gaps and drifting particles, and will not mistake fixed stains or the structural texture of the roller shutter door itself for changes in the alarm situation.

[0029] When obtaining the infrared temperature difference, the controller maps the synchronization marker to the acquisition timeline of the infrared thermal imaging module and selects the infrared image corresponding to the synchronization marker. It should be noted that the infrared temperature difference does not require a single robot to penetrate the metal roller shutter door to directly detect the front and back sides of the door. Instead, it collects data within the visible area of ​​the door gap between the lower edge of the roller shutter door and the ground. Using the airflow direction of the door gap as a reference, the controller sets temperature sampling zones in the adjacent areas on the windward and leeward sides of the door gap opening contour. Both temperature sampling zones avoid the metal door body, areas with obvious water accumulation and reflection, and areas obscured by the robot body, prioritizing coverage of the adjacent ground area below the door gap that can be directly observed by the infrared thermal imaging module.

[0030] For example, for infrared temperature difference, temperature sampling bands are set up in the adjacent areas on the windward and leeward sides of the door gap, respectively. The representative temperature of the effective pixels in the windward temperature sampling band is denoted as... The representative temperature of the effective pixels within the leeward side temperature sampling band is denoted as... Then infrared temperature difference Determined in the following manner: The temperature represented is the median value of the effective pixel temperature within the temperature sampling band. An effective pixel refers to a pixel that has not fallen into the metal door area, the area with obvious water accumulation and reflection, or the area obscured by the robot body, and has persisted for at least one complete landing cycle. When the infrared temperature difference is positive, it indicates that the temperature in the adjacent area on the leeward side is higher; when the infrared temperature difference is negative, it indicates that the temperature in the adjacent area on the windward side is higher. A change in infrared temperature difference direction from positive to negative or vice versa is recorded as a change in infrared temperature difference direction.

[0031] The controller extracts stable temperature data that persists for at least one landing cycle within two temperature sampling bands, and obtains the infrared temperature difference based on the representative temperature difference between the two sampling bands. For temperature abrupt changes lasting less than one landing cycle, the controller does not directly use them to calculate the infrared temperature difference, but retains them as a sudden risk marker. This reduces errors caused by water droplets adhering to the lens, metal thermal reflection, or water reflection, while ensuring that short-term localized temperature rises occurring in the early stages of battery thermal runaway are not missed.

[0032] Upon obtaining the acoustic echo delay, the controller selects a time window with weaker chassis vibration between two adjacent synchronization markers and transmits coded acoustic short pulses through the speaker. These coded acoustic short pulses consist of multiple short segments of varying durations arranged in a fixed order. After receiving the reflected signal, the microphone array first filters out exhaust fan noise, alarm sounds, and motor noise according to the coded order, and then selects effective reflections aligned with the door gap direction within the preset echo time window.

[0033] The controller identifies the primary reflection formed by the lower edge of the roller shutter door or the edge of the door seam based on the receiving direction of the microphone array and the echo arrival time, and also identifies the secondary reflection that returns after passing through the ground. For reflection combinations that can be repeated by two consecutive coded acoustic short pulses, the controller records the arrival order of the primary and secondary reflections and the time interval between them to obtain the acoustic echo delay. For example, for the acoustic echo delay, the arrival time of the primary reflection from the edge of the door seam is recorded as... The arrival time of the second reflection from the ground is recorded as Then the acoustic echo time delay Determined in the following manner: The controller records the arrival order of the first and second reflections while recording the acoustic echo delay. When the coded acoustic short pulse is transmitted twice consecutively, and both transmissions result in the same arrival order, this arrival order is determined as the valid arrival order. Echo combinations that occur only once are not used to form the gap reference interval.

[0034] If the reflection combination occurs only once, or its encoding order cannot correspond to the transmission sequence, it is ignored as a spray sound or a temporary collision sound.

[0035] When obtaining the rate of change in gas concentration, the robot first determines the airflow direction through the door gap based on the operating direction of the exhaust fan and the continuous drift direction of light particles in the visible light image. When the rated direction of the exhaust fan is inconsistent with the actual drift direction of the particles, the drift direction of the particles that remains stable over multiple consecutive landing cycles is taken as the airflow direction through the door gap. The robot then moves to the windward side of the door gap and maintains a preset distance between the gas sensor and the door gap.

[0036] To form a reproducible gas sensor recovery segment, the robot dog does not simply stop stationary. Instead, it first performs a short lateral sweep along a direction intersecting the airflow direction at the door gap, allowing the gas sensor to sequentially pass through the edge of the door gap plume and the relatively stable region on the windward side. After completing the lateral sweep, the robot dog maintains its orientation and pauses briefly. The controller defines the continuous data segment from the gas sensor's plume-disturbed state to a stable output state as the recovery segment. Based on the continuous trend of gas concentration change within the recovery segment, the gas concentration change rate is obtained, and the gas concentration corresponding to the starting point of the gas sensor recovery segment is recorded as... The gas concentration corresponding to the end of the recovery segment is denoted as The duration of the recovery segment is recorded as Then the rate of change of gas concentration Determined in the following manner: The recovery phase ends when the concentration changes corresponding to two consecutive synchronization markers on the gas sensor do not exceed a stable threshold. This stable threshold is the larger of the nominal repeatability error of the gas sensor and the fluctuation range of the concentration samples taken five consecutive times before the lateral sweep. The collaborative control unit simultaneously records the number of synchronization markers crossed during the recovery phase. An increase in the number of synchronization markers crossed during the recovery phase indicates an increase in the degree of local plume congestion.

[0037] For concentration spikes that last only a very short time and cannot continue into the stagnant phase, they are retained as transient markers but not directly used as the rate of change of gas concentration.

[0038] Finally, using the synchronization marker as an index, the controller arranges the visible light occlusion rate, infrared temperature difference, acoustic echo delay, gas concentration change rate, sudden risk marker, transient marker, and corresponding pose information obtained within the same landing cycle window into a set of observation records. Each set of observation records is arranged consecutively according to the robot dog's movement sequence, forming an alarm observation sequence. Different robot dogs use the same data fields and arrangement order, and upload their respective alarm observation sequences to the collaborative control terminal.

[0039] Understandably, when the controller generates each set of observation records, it uses the contour position marker corresponding to the intermittent blanking unit as the associated field of visible light occlusion rate, the side of the two temperature sampling bands and the direction of infrared temperature difference as the associated field of infrared temperature difference, and the arrival order of the first reflection and the second reflection as the associated field of acoustic echo delay, and writes it into the alarm observation sequence.

[0040] Please see Figure 2 As shown, based on the alarm observation sequence and the airflow direction of the roller shutter door gap, the modal inversion boundary where the response direction of different sensing modes reverses is determined, and the alarm confidence vectors on both sides of the boundary are obtained.

[0041] After obtaining the alarm observation sequence, the collaborative control terminal identifies the modal inversion area near the roller shutter door gap caused by the combined effects of airflow deflection, water mist migration, ground water reflection, and metal door reflection. The modal inversion area refers to a local area where the alarm source has not undergone a substantial change in location, but at least two sensing modalities present opposite alarm directions in the vicinity of the door gap.

[0042] It should be noted that the collaborative control terminal is a computing node with functions of data reception, data association, path analysis, and task distribution. It is used to receive alarm observation sequences uploaded by each robot dog, process the observation records of different robot dogs in a unified manner, and send detection crossing, path adjustment, and collaborative response commands to each robot dog. The collaborative control terminal can be set up in a security control center server, an edge computing device in an underground parking garage, or any local controller of a robot dog with collaborative computing capabilities.

[0043] During the aforementioned alarm observation sequence formation process, each robot dog has already obtained the initial airflow direction at the door gap based on the working direction of the exhaust fan and the drift direction of light particulate matter. The collaborative control terminal further uses the gas concentration change rate of the adjacent areas on the windward and leeward sides to correct the initial airflow direction at the door gap. Specifically, when light particulate matter continuously leaves the door gap and the gas sensor recovery segment in the corresponding area lasts for a relatively long time, the side where that area is located is determined as the leeward side; the side where the particulate matter converges towards the door gap and the gas sensor recovery segment returns to a stable state relatively quickly is determined as the windward side. This yields the corrected airflow direction at the door gap. The airflow direction at the roller shutter door gap referred to thereafter refers to the corrected airflow direction at the door gap. A relatively quick return to a stable state means that the number of synchronization markers crossed by the recovery segment is small, and there is at least a difference of 1 synchronization marker.

[0044] Based on the pose information in each set of observation records, the collaborative control unit establishes a folded sampling zone on the environmental map. This folded sampling zone is a virtual projection area used to collect observation records, rather than a fixed trajectory that the robot dog must actually traverse. The folded sampling zone includes a windward branch, a branch near the door gap, and a leeward branch. The windward and leeward branches extend along the airflow direction of the roller shutter door gap, while the branch near the door gap connects the two aforementioned branches near the lower edge projection area of ​​the roller shutter door, creating a folded shape around the area affected by the airflow through the door gap.

[0045] The width of the folding sampling zone is determined based on the robot dog's pose positioning error and the lateral sway range within a landing cycle, avoiding the roller shutter door guide rail, fixed lighting, and areas where water accumulates over a long period. For any observation record, the collaborative control terminal projects it onto the nearest folding sampling zone branch based on the corresponding pose information, and records the distance between the observation record and the projected position of the lower edge of the door gap, as well as the arrangement order along the airflow direction of the door gap.

[0046] The collaborative control unit then pairs the observation records from the windward and leeward branches. Each pair of records should meet the following requirements: the distance between the record and the projected position of the lower edge of the door gap is within a preset allowable range; the projected positions along the airflow direction of the door gap correspond to each other; and the acquisition time difference does not exceed the time window during which the airflow state of the door gap remains stable. The preset allowable range is the sum of the robot dog's maximum pose positioning error and the low-speed movement distance within one landing cycle.

[0047] During pairing, observation records obtained by the same robot dog covering both the windward and leeward branches at different inspection phases are prioritized. If the same robot dog does not cover both branches, alarm observation sequences generated by another robot dog in adjacent time periods are used to supplement the records. The supplementary records should meet the following requirements: synchronization marker interval, low-speed gait, and fuselage orientation should be within preset allowable ranges. This reduces the impact of gait differences between different robot dogs and fuselage vibration on cross-side comparison results.

[0048] The collaborative control unit first establishes a gap-penetration reference based on the acoustic echo time delay. It should be noted that this gap-penetration reference does not mean the robot dog passes through the gap in the roller shutter door, but rather that the reflection path of the coded acoustic short pulse reaches a stable transition after passing through the area affected by the gap. Specifically, the collaborative control unit reads the correlation field of the acoustic echo time delay to identify the arrival order of the primary reflection at the edge of the gap and the secondary reflection from the ground. As the observation position corresponding to the robot dog gradually approaches the area affected by the airflow through the gap, if the arrival order of the two types of reflections stably switches in the observation records corresponding to three consecutive synchronization markers, then the consecutive position interval where the stable switch occurs is recorded as the gap-penetration reference interval.

[0049] If the arrival sequence changes only at a single synchronization marker, or fails to be maintained in subsequent observation records after the change, it is considered an occasional acoustic anomaly caused by water droplet impact, temporary obstacle movement, or ground water disturbance, and no gap reference interval is formed. If no gap reference interval is formed in the current alarm observation sequence, the collaborative control terminal continues to control the robot dog to supplement sampling, and the modal inversion boundary is temporarily uncertain.

[0050] After establishing the reference interval for the slit, the collaborative control unit uses this reference interval as the center and compares the paired observation records segment by segment along the folded sampling band. For visible light occlusion rate, it reads the intermittent blanking unit contour position marker in its associated field and determines whether the intermittent blanking unit continuously migrates from the windward branch to the leeward branch, or vice versa. For infrared temperature difference, it reads the temperature sampling bands of the adjacent areas on the windward and leeward sides and determines whether the direction of the infrared temperature difference changes from higher on the windward side to higher on the leeward side, or vice versa. For gas concentration change rate, it determines whether the gas sensor recovery segment changes from a rapid return to a stable state to a slow return to a stable state, or vice versa.

[0051] The reversal of response directions for different sensing modes does not need to occur at the same synchronization marker, but both should fall within a continuous observation window set around the slit reference interval. This continuous observation window covers the slit reference interval and extends at least one foot cycle before and after it. This allows for the accommodation of time lags existing between water mist migration, heat diffusion, and the gas sensor recovery process.

[0052] In this invention, when at least two of the sensing modes—visible light occlusion rate, infrared temperature difference, and gas concentration change rate—show a reversal in response direction within the continuous observation window, and the slit-crossing reference interval corresponding to the acoustic echo delay persists, the collaborative control terminal marks the corresponding continuous observation window as a candidate inversion segment. For cases where only one sensing mode changes, or where only one discontinuous acoustic sequence switch occurs, the collaborative control terminal does not generate a candidate inversion segment.

[0053] The collaborative control unit selects one robot dog as the verification robot dog from those currently not performing other high-priority alarm tasks, whose remaining battery power meets the round-trip sampling requirements, and which are located within a safe passage area. The collaborative control unit controls the verification robot dog to perform lateral round-trip maneuvers within the passable area outside the roller shutter door, along a verification path intersecting the airflow direction of the roller shutter door gap. The verification path passes through the candidate inverted section, but the verification robot dog is not required to enter below the roller shutter door or cross the roller shutter door gap.

[0054] The robot dog was verified to maintain the same low-speed gait on both the outward and return journeys, and to continue using the structural vibrations generated by its forefoot landings as synchronization markers. The collaborative control unit interleaved and stitched the outward and return journey observation records according to these synchronization markers. Specifically, the observation record corresponding to a certain spatial position on the outward journey was arranged as an adjacent record with the observation record corresponding to the same or adjacent spatial position on the return journey. Then, it was determined whether the same type of response direction reversal occurred when passing that position twice.

[0055] For candidate inversion segments that occur only on the outward or return journey, they are identified as unidirectional anomalies caused by transient water mist, temporary obstruction, mobile lighting, or aircraft turning, and are therefore eliminated. For candidate inversion segments where the response direction is reversed throughout the round trip, and the spatial offset of the seam reference interval does not exceed the movement distance corresponding to one landing cycle, the collaborative control terminal extracts the location point with the most concentrated response direction reversal within the candidate inversion segment and connects the corresponding location points in different sampling paths to form a modal inversion boundary. The modal inversion boundary divides the folded sampling zone into two local areas with different alarm response directions.

[0056] The collaborative control unit collects the staggered observation records from both sides of the modal inversion boundary and generates an alarm confidence vector according to a fixed order of visible light sensing, infrared sensing, acoustic sensing, and gas sensing. Each sensing mode corresponds to a confidence term, and each confidence term records at least the response direction, persistence state, and confidence level. The response direction indicates which side of the modal inversion boundary the corresponding sensing mode points to; the persistence state indicates whether the response direction persists during adjacent synchronization markers and round-trip sampling.

[0057] A perception result that occurs in both the outward and return journeys, remains stable across three consecutive synchronization markers, and exhibits a consistent alert direction with at least one other perception modality is classified as having a high confidence level. A perception result that occurs in both the outward and return journeys but remains stable only at local synchronization markers, or has not yet formed a consistent alert direction with other perception modalities, is classified as having a medium confidence level. A perception result that occurs only briefly during unidirectional movement, or cannot be repeated at adjacent synchronization markers, is classified as having a low confidence level.

[0058] The collaborative control unit further writes the persistent state of the modal inversion boundary into the alarm confidence vector. For modal inversion boundaries that remain stable after round-trip sampling, the retention priority of high-confidence terms in the alarm confidence vectors on both sides of the boundary is increased; for modal inversion boundaries that continuously drift in spatial location or are supported only by a single sensing modality, the calling priority of the corresponding alarm confidence vector is reduced, and supplementary sampling is triggered.

[0059] Therefore, the alarm confidence vectors on both sides of the boundary not only retain the response direction of each sensing mode, but also record the acoustic seam reference, round-trip verification results, and the continuous state of response direction reversal. When building the alarm shadow map later, the collaborative control terminal can write the alarm confidence vector into the corresponding node, avoiding multiple robot dogs from heading to the wrong side of the roller shutter door based solely on a single temperature peak or a single gas concentration peak.

[0060] An alarm shadow map is established based on the alarm confidence vector and the passable path. The alarm shadow map includes shadow nodes that record risk values ​​and information missingness, as well as occupancy gates that record the cost of channel occupancy.

[0061] After obtaining the alarm confidence vectors on both sides of the modal inversion boundary, the collaborative control terminal establishes an alarm shadow map based on the alarm confidence vectors and the passable paths in the underground parking garage. The alarm shadow map is used to describe which obstructed areas may still hide the real alarm source when modal inversion exists near the roller shutter door, which areas have not been effectively verified, and which narrow paths are likely to cause response blockage when multiple robot dogs enter at the same time.

[0062] In this embodiment, the passable path is not solely based on pre-marked roads in the environmental map, but is jointly updated by the collaborative control terminal based on the environmental map, the robot dog's real-time pose information, and temporary obstacles identified by the visible light camera module. For path segments near the lower edge of roller shutter doors, corner bends, roller shutter door guide rails, vehicle sides, vehicle bottoms, and fire-fighting facility bases, the collaborative control terminal records their passable width, ground flatness, and spatial relationship with the modal inverted boundary. Areas where the water depth exceeds the robot dog's permissible wading height, are blocked by falling objects, or are completely enclosed by parked vehicles are not considered passable paths; areas where the robot dog can pass alone but cannot meet in parallel are retained as restricted passable paths.

[0063] The collaborative control unit first projects the alarm confidence vectors on both sides of the modal inverted boundary onto the corresponding passable paths. Each alarm confidence vector contains the response direction, duration, and confidence level of visible light sensing, infrared sensing, acoustic sensing, and gas sensing. The collaborative control unit reads the sensing items with high or medium confidence levels and uses the direction of that sensing item pointing to one side of the modal inverted boundary as the initial alarm direction corresponding to that sensing item.

[0064] Due to water vapor migration, heat diffusion, airflow deflection, and acoustic reflection near the door gap, the response direction of a single sensing mode may not directly point to the actual alarm source. To avoid mistaking the apparent direction after mode inversion for the actual alarm source direction, the collaborative control terminal performs reverse backtracking along the passable path in the opposite direction of its initial alarm direction for each sensing item with a high or medium confidence level. The reverse backtracking paths formed by different sensing items can be the same, or they can enter different obstructed areas due to different sensing mechanisms.

[0065] The reverse backtracking is not a simple extension of a straight path, but rather unfolds segment by segment along the actually traversable path that the robot dog can reach. When the backtracking path encounters a corner, roller shutter door rail, vehicle side, vehicle bottom, or fire-fighting facility base, the collaborative control terminal determines whether an observation interruption or directional shift is likely to occur at that location based on the perception modality type. Specifically, visible light perception corresponds to physical obstruction and water mist obstruction, infrared perception corresponds to metal heat reflection and ground water reflection, acoustic perception corresponds to door seam edge reflection and ground secondary reflection, and gas perception corresponds to airflow around door seams and stagnation in vehicle bottom cavities. If the backtracking direction reverses at the above-mentioned locations, or if a certain perception item cannot be further verified after that location, a shadow node is inserted at that location.

[0066] For multiple shadow nodes formed by different sensing items, if they are located at the entrance of the same obstruction structure, the same angle of the roller shutter door guide rail, the same vehicle bottom entrance, or the same corner bend, the collaborative control terminal merges them into a composite shadow node, and retains the source and initial alarm direction of each sensing item in the composite shadow node. For multiple shadow nodes that are spatially separated but led out by the same modal inverted boundary, they are retained separately. Multiple shadow nodes and the passable paths between them constitute shadow branches. Shadow branches not only cover the areas that the robot dog has already traversed, but also cover local areas that have not yet been directly observed due to corner bends, guide rail obstructions, or vehicle bottom cavities.

[0067] For example, when infrared sensing points to the outside of the roller shutter door guide rail while gas sensing points to the inside, the collaborative control unit does not immediately discard either direction. Instead, it traces back along the reverse path of infrared sensing to the angle between the roller shutter door guide rail and the wall, and then traces back along the reverse path of gas sensing to the bottom entrance of the vehicle through which the airflow passes in the door gap. It then inserts or merges the corresponding shadow nodes. For shadow nodes directly connected to the modal inversion boundary and capable of serving as the starting point for supplementary detection in the obstructed area, the collaborative control unit further marks them as boundary nodes. When selecting subsequent detection crossing locations, priority is given to filtering from these boundary nodes.

[0068] For each shadow node, the collaborative control unit determines the upstream and downstream shadow nodes according to the reverse backtracking direction. The side closer to the modal inversion boundary and with more effective observation records is designated as the upstream, while the side continuing deeper into the occluded area and without completed perception verification is designated as the downstream. For composite shadow nodes formed by the merging of multiple perception items, the collaborative control unit determines the upstream and downstream for each perception item separately and records them separately within the same node.

[0069] The collaborative control unit aggregates high-confidence sensing items from upstream shadow nodes and checks whether these sensing items can still be verified by the corresponding sensing mode after the path detour through the current shadow node. For example, infrared sensing maintains high confidence in front of the roller shutter door guide rail, but cannot continue to obtain effective temperature sampling in the vehicle bottom area behind the guide rail; acoustic sensing can bypass the guide rail and continue to form echoes, but visible light sensing cannot identify intermittent blanking units due to vehicle body obstruction. In this case, the collaborative control unit records infrared sensing and visible light sensing as downstream missing items, respectively, and retains the acoustic sensing items that can still be verified. The missing items are used to characterize that a certain sensing mode is not judged to be risk-free, but rather that sufficient observations have not been obtained due to obstruction, heat reflection, missing acoustic paths, or airflow detours.

[0070] The collaborative control unit converts the landing cycles of consecutively traversed missing items into continuous unchecked segments along the shadow branch based on the pose information corresponding to the synchronization markers, and writes the information missing degree accordingly. Specifically, when a missing item corresponds to only one landing cycle, and a valid observation can be obtained again at an adjacent shadow node or adjacent path segment, the missing item is recorded as a short-term missing item, and the information missing degree of the corresponding node is set to low. If a missing item corresponds to multiple consecutive landing cycles, and no corresponding sensing modality check is obtained after extending along the shadow branch to the downstream shadow node, it is recorded as a persistent missing item, and the information missing degree of the corresponding node is set to medium. If at least two sensing modalities simultaneously form a persistent missing item, and there is still a passable path or a restricted passable path downstream, the information missing degree of the corresponding node is set to high. The collaborative control unit also writes risk values ​​based on the retention status of upstream high-confidence items before and after path turning. If a high-confidence item persists before entering a shadow node, and after path turning, the same warning direction is maintained by the same sensing modality or another sensing modality, the risk value of the shadow node is set to high. If a high-confidence term temporarily disappears after a path turn, but adjacent upstream shadow nodes still retain anomalies in the same direction of infrared temperature difference, gas concentration change rate, or acoustic echo delay, then the risk value of that shadow node is set to medium and it is marked as a risk node to be reviewed. If multiple sensing modalities do not remain abnormal after round-trip verification, and there are no persistent missing terms downstream, then the risk value of that shadow node is set to low.

[0071] In this invention, risk value and information deficiency are recorded separately and are not interchangeable. For example, the area under a vehicle may have a high information deficiency due to the simultaneous failure of visible light and infrared sensing, but its risk value has not yet reached a high level; the angle of the roller shutter door guide rail may also have a high risk value because the infrared temperature difference and gas concentration change rate continuously point to the same side, even if its information deficiency is low. By adopting the above distinction method, the collaborative control terminal can avoid selecting a path based solely on the existing alarm intensity, while ignoring narrow, obstructed areas that have not yet been effectively detected.

[0072] The collaborative control unit projects the real-time pose information and synchronization markers of each robot dog onto the shadow branch. For any path segment, the collaborative control unit determines whether the path segment can accommodate two robot dogs simultaneously based on the channel width in the environmental map, the lateral occupancy width of the robot dog, and the lateral swing range within a landing cycle. If a path segment cannot accommodate two robot dogs passing in parallel within a single landing cycle, or if two robot dogs cannot safely meet in the path segment, then the path segment is marked as an occupied gate. The occupied gate is a restricted path segment written into the alarm shadow map in the form of path constraint nodes, and its entrance and exit are still connected to adjacent passable paths, respectively.

[0073] The occupancy gate can correspond to narrow passages between roller shutter door tracks and walls, gaps between parked vehicles and fire-fighting equipment, vehicle bottom entrances, and one-way detour sections formed by the edge of accumulated water. For each occupancy gate, the collaborative control terminal records its entrance position, exit position, estimated number of landing cycles required for passage, current occupancy status, and the estimated arrival order of adjacent robot dogs. Based on the pose information, movement order, and synchronization mark interval of each robot dog, the collaborative control terminal determines the expected order of entry and exit from the occupancy gate and writes the channel occupancy cost.

[0074] Specifically, when a certain gate is not yet used by other robot dogs, and the currently scheduled robot dog can pass through within a preset number of landing cycles without blocking the exit path of other robot dogs, its channel occupancy cost is set to low.

[0075] It should be noted that the preset quantity represents the estimated number of landing cycles required to pass through, determined based on the length of the occupied door and the low-speed movement distance of the robot dog within one landing cycle.

[0076] When another robot dog has already entered the occupied gate, or when two robot dogs are expected to arrive at the occupied gate consecutively from the same direction within adjacent landing cycles, set its channel occupancy cost to medium. When two robot dogs are expected to arrive at the occupied gate from opposite directions within adjacent landing cycles, or when the occupied gate is located between a high-risk node and a high-information-deficient node and its blockage would cut off the subsequent evacuation path, set its channel occupancy cost to high and restrict multiple robot dogs from entering simultaneously.

[0077] The collaborative control unit uses shadow nodes and occupancy gates as graph nodes and passable paths as graph connections to establish an alarm shadow graph. Occupancy gates are written into the graph structure as path constraint nodes, and shadow nodes are written into the graph structure as perception status nodes. Each shadow node records at least its location, node type, risk value, information missing degree, upstream shadow nodes, downstream shadow nodes, associated perception items, and connection relationships with modal inverted boundaries. Each occupancy gate records at least its entrance location, exit location, estimated passage time, current occupancy status, and channel occupancy cost. For boundary nodes, their corresponding modal inverted boundaries and accessible shadow branches are further recorded.

[0078] In the warning shadow map, high-risk nodes, nodes with high information gaps, and occupancy gates with high channel occupancy costs can coexist, but they represent different meanings. High-risk nodes indicate that multimodal sensing results are continuously pointing to the area; nodes with high information gaps indicate that there are still sensing gaps within the obscured area that have not yet been verified; and high channel occupancy costs indicate that multiple robot dogs entering simultaneously are likely to cause congestion. Based on this, the collaborative control unit can prioritize boundary nodes with high information gaps but still safe passage conditions in subsequent steps, controlling at least one robot dog to conduct detection and crossing, and preventing multiple robot dogs from disorderly gathering in narrow channels.

[0079] It should be noted that the incident shadow map links the incident confidence vectors on both sides of the modal inverted boundary, the perception gaps within the occluded area, and the restricted access conditions in the underground parking garage into the same graph structure. Compared to responding solely based on the shortest path or the confidence of a single incident, this processing method can identify local incident shadows formed by corners, guide rails, and the bottom of vehicles, and provides a traceable data foundation for subsequent correction of the incident shadow map and generation of off-peak response paths.

[0080] Based on the aforementioned alarm shadow map, the boundary node with the largest information missing degree is selected, and at least one robot dog is controlled to perform a detection crossing. The alarm shadow map is then corrected using the observation difference before and after the crossing to obtain the target alarm area.

[0081] After establishing the alarm shadow map, the "detection crossing" refers to the robot dog entering the corresponding shadow branch from the boundary node and crossing the obstruction and turning points formed by corners, roller shutter door rails, vehicle sides, vehicle bottom entrances, or fire-fighting facility bases. Detection crossing does not require the robot dog to pass through roller shutter door gaps or enter areas that have been determined to be impassable.

[0082] The collaborative control terminal first updates the information missing degree of each boundary node. For each boundary node, the collaborative control terminal reads the shadow nodes downstream segment by segment along its corresponding shadow branch, stopping at impassable paths or occupancy gates with high channel occupancy costs. The highest information missing level among the reachable downstream shadow nodes is written to the corresponding boundary node. When multiple downstream shadow nodes have the same highest information missing level, the number of consecutive landing cycles of the missing items corresponding to that level is further accumulated and used as the continuous missing range of the boundary node.

[0083] The collaborative control unit compares the information missing degree of each boundary node in high, medium, and low order, and determines the boundary node with the highest information missing degree as the candidate detection crossing entrance. When multiple boundary nodes have the same information missing degree, the boundary node with the larger continuous missing range is selected first. When the continuous missing range is still the same, the boundary node with a large number of downstream high-risk shadow nodes and whose corresponding channel occupancy cost does not reach a high level is selected first as the detection crossing entrance.

[0084] If all candidate detection crossing entrances are associated with occupied gates with high channel occupancy costs, the collaborative control terminal will suspend issuing entry commands and update the alarm shadow map based on the real-time pose information of existing robot dogs after exiting the occupied gates. If multiple robot dogs need to be called simultaneously, different robot dogs will prioritize entering shadow branches that do not share the same occupied gate; when they must share the same occupied gate, they will enter sequentially according to the expected arrival order recorded by the synchronization marker to avoid forming opposing blockages within the restricted passable path.

[0085] The collaborative control unit selects at least one robot dog as a detection robot dog from among those located in a safe passage area, with sufficient remaining power for round trips, and not currently performing other high-priority alarm tasks. Sufficient remaining power for round trips means that the robot dog can complete the journey from its current location to the detection crossing entrance, from the detection crossing entrance to the planned detection endpoint, and back to the safe passage area from the planned detection endpoint, while retaining sufficient remaining power for emergency evacuation.

[0086] The collaborative control unit determines the obstruction and turning point based on the location where the missing items in the alarm shadow map begin to appear consecutively and the location where the perceived items shift direction. The detection robot sets one stopping point before, during, and after passing the obstruction and turning point. If the obstruction and turning point itself does not provide stable standing conditions, the nearest passable path location that allows the robot to maintain stable foot support is selected as the corresponding stopping point, and the positional offset between this stopping point and the obstruction and turning point is recorded.

[0087] At each stop, the detection robot first performs a short-range lateral sweep along the direction intersecting with the airflow projection from the roller shutter door gap, allowing the gas sensor to sequentially pass through the edge of the local plume and a relatively stable region; then, maintaining its orientation, it pauses briefly. Each brief pause covers at least three consecutive synchronization markers to create at least two time windows of relatively weak body vibration between adjacent synchronization markers, and each pause completes two coded acoustic short pulse transmissions.

[0088] The detection robot collects visible light occlusion rate, infrared temperature difference, acoustic echo delay, gas concentration change rate, and corresponding pose information at each stopping point. For the visible light occlusion rate, the contour position markers of the intermittent blanking units are retained. For the gas concentration change rate, the number of synchronization markers crossed by the gas sensor recovery segment and the recovery direction are retained.

[0089] For infrared temperature difference, when the visible area of ​​the roller shutter door gap is still within the effective imaging range of the infrared thermal imaging module, temperature sampling bands are continued to be used on the adjacent areas of the windward and leeward sides of the door gap. When the roller shutter door gap is blocked by the side of a vehicle, a corner of a wall, or the roller shutter door guide rail, the collaborative control terminal sets local temperature sampling bands on the upstream and downstream sides of the obstruction and reversal position, and records the reference source of the local temperature sampling bands. The local temperature sampling bands are only used for local verification before and after the obstruction and reversal position, and are not directly mixed and compared with the infrared temperature difference formed by the door gap temperature sampling bands.

[0090] For acoustic echo delay, the aforementioned door gap acoustic reference is continued to be used as long as the primary reflection from the door gap edge and the secondary reflection from the ground can still be effectively identified. When the door gap acoustic reference becomes unavailable due to obstruction, the cooperative control unit identifies the primary reflection formed by the edge of the obstruction's turning position and the secondary reflection returning from the ground, and records them as local acoustic references. The source marker of the local acoustic reference is written into the observation record along with the acoustic echo delay. Only acoustic echo delays with the same reference source are compared before and after.

[0091] If a certain sensing mode cannot obtain an observation record that meets the validity conditions at the dwell point, the cooperative control end will continue to retain it as a missing item and will not directly interpret invalid observations as low risk. For example, when the bottom of the vehicle completely obscures the visible light outline, no direction judgment of visible light obscuration rate will be generated; when the coded acoustic short pulses cannot form a repeatable reflection combination, no acoustic reflection arrival order judgment will be generated.

[0092] The collaborative control unit arranges the observation records from the three stopping points in a reverse staggered manner according to the synchronization markers. Specifically, taking the occlusion turn position as a reference, it first reads the downstream observation records after the occlusion turn position, then reads the intermediate observation records when passing through the occlusion turn position, and finally reads the upstream observation records before the occlusion turn position. For multiple observation records within each stopping point, they are arranged from near to far according to the distance between the corresponding pose information and the occlusion turn position. This forms a sequence of observation records that are checked back from downstream to upstream.

[0093] The reverse staggered arrangement is used to accommodate the different response speeds of various sensing modalities. Visible light occlusion rate and acoustic echo delay can usually respond relatively quickly to changes in occlusion and deflection, while infrared temperature difference is affected by heat diffusion and has a certain delay, and gas concentration change rate is affected by local airflow stagnation and has a more significant recovery delay. By tracing back from downstream to upstream, it is possible to avoid mistaking gas sensing lag for new alarm sources based solely on the order of acquisition time.

[0094] The collaborative control unit then extracts the observation difference. For visible light sensing, when an intermittent blanking unit, after entering the occlusion turning position, no longer continues to migrate downstream but reappears in the continuous observation record near the upstream contour position, it is recorded as an intermittent blanking unit position retreat. For infrared sensing, when the infrared temperature difference from the same reference source changes from being higher in the windward adjacent area to being higher in the leeward adjacent area, or from being higher in the upstream local temperature sampling band to being higher in the downstream local temperature sampling band, it is recorded as an infrared temperature difference direction switch.

[0095] For acoustic sensing, an acoustic reflection arrival order reset is defined as follows: when the arrival order of primary and secondary reflections from the same reference source returns to the stable arrival order before the obstruction / turning position after passing the obstruction / turning position, and this reset state is repeated in two consecutive coded acoustic short pulses. For gas sensing, a gas recovery segment extension is defined as follows: when the recovery segment after the obstruction / turning position spans at least one more synchronization marker than the recovery segment before the obstruction / turning position, and this extension state is repeated at the downstream dwell point.

[0096] The collaborative control unit establishes correction windows segment by segment along the shadowed branch. Each correction window covers no more than two adjacent landing cycles, and its spatial range does not exceed the travel distance of the detection robot within two landing cycles. When at least two of the following occur within the same correction window: intermittent blanking unit position retraction, infrared temperature difference direction switching, acoustic reflection arrival order reset, and gas recovery segment extension, a correction node is inserted at the corresponding position of that correction window. If different observation differences occur in multiple spatially separated correction windows, correction nodes are inserted separately, and they are not merged.

[0097] The collaborative control unit re-examines the missing items in the corresponding shadow branch based on the correction node. When the detection robot dog re-acquires a valid observation record of a certain perception mode downstream of the correction node, and this valid observation record appears repeatedly in two consecutive synchronization markers, the perception mode is updated from a missing state to a verified state and removed from the consecutive unverified segments. The removal process involves reclaiming verified missing items.

[0098] After the missing items are recovered, the collaborative control terminal re-determines the information missing degree of the corresponding shadow node based on the number of consecutive landing cycles traversed by the remaining missing items. If the missing item corresponds to only one landing cycle and a valid observation can be obtained again in an adjacent shadow node, the information missing degree is updated to low. If the missing item corresponds to multiple consecutive landing cycles and persists in downstream shadow nodes, the information missing degree is updated to medium. If at least two sensing modes are simultaneously and persistently missing, and a passable path still exists downstream, the information missing degree is updated to high.

[0099] The collaborative control terminal simultaneously updates risk values. When at least two sensing modalities downstream of the correction node maintain the same alarm indication across three consecutive synchronization markers, and at least one of the sensing modalities has a high confidence level, the risk value of the corresponding shadow node is updated to high. When only one sensing modality remains abnormal continuously, or when the abnormal directions of multiple sensing modalities are not yet consistent, the risk value of the corresponding shadow node is updated to medium, and it is marked as a risk node awaiting review. When the missing item has been recovered, and there are no continuously maintained multimodal anomalies downstream, the risk value of the corresponding shadow node is updated to low.

[0100] Finally, the collaborative control unit starts from the correction node and connects continuously high-risk shadow nodes along the shadow branch. When adjacent shadow nodes maintain the same alarm direction in three consecutive synchronization markers, and there are no unverified high-information missing nodes or high-cost occupancy gates between them, they are classified into the same target alarm area. If there is only one high-risk shadow node, then that shadow node and its adjacent verified path segment are jointly determined as the target alarm area. If there are multiple separate sets of high-risk shadow nodes, then multiple candidate target alarm areas are formed for subsequent collaborative response steps to sort and process according to risk value and channel occupancy cost.

[0101] Please see Figure 3 As shown, based on the revised alarm shadow map, each robot dog is divided into detection, control and communication relay roles, and staggered response paths are generated according to the channel occupancy cost.

[0102] After obtaining the corrected alarm shadow map and the target alarm area, the roles of each robot dog participating in the response are assigned, and staggered response paths are generated. The staggered response paths are used to avoid multiple robot dogs simultaneously entering the gates formed by roller shutter guide rails, vehicle gaps, or water accumulation bypass sections, and to prevent detection robot dogs and control robot dogs from blocking each other in narrow paths.

[0103] The collaborative control terminal extracts high-risk shadow nodes, boundary nodes, and occupied gates connected to the target alarm area from the corrected alarm shadow map, and determines the path that can return from the target alarm area to the safe passage area as the evacuation path. Along the evacuation direction, the collaborative control terminal sequentially checks each occupied gate outward from the gate closest to the target alarm area to form the order of occupied gates.

[0104] For each occupied gate, the collaborative control unit reads its entrance position, exit position, channel occupancy cost, and the estimated number of landing cycles required for passage. The estimated number of landing cycles required for passage is determined based on the path length from the entrance to the exit of the occupied gate, the corresponding low-speed gait of the robot dog, and the range of lateral swing. The collaborative control unit uses the synchronization mark corresponding to when the robot dog is expected to reach the entrance of the occupied gate as the starting point, extends the required number of landing cycles, and writes the corresponding extended synchronization mark to the occupied gate as an exit synchronization mark.

[0105] The collaborative control unit assigns roles to the robot dog based on its current location, remaining battery power, accessibility to shadowed tributaries, and communication link status. Robot dogs capable of entering the downstream shadowed tributary of the target alert area through the occupied gate, and whose remaining battery power is sufficient to cover the entry, detection, and withdrawal processes, are assigned the detection role. The detection role continues to collect visible light occlusion rate, infrared temperature difference, acoustic echo delay, and gas concentration change rate, and supplements observation records near the target alert area.

[0106] Robot dogs capable of reaching the boundary node and not occupying the evacuation path or obstructing the entrance to the occupation gate after stopping are designated as control roles. Control roles remain within the boundary node or a nearby passable path to prevent other robot dogs from accidentally entering high-cost occupation gates and to preserve the exit path when the detection role evacuates. If multiple shadowed paths exist near the boundary node, control roles are prioritized for positions that can simultaneously cover the entrances of multiple shadowed paths.

[0107] Robot dogs located near path turning points that can maintain bidirectional communication with both the collaborative control unit and at least one detection robot dog are classified as communication relay roles. The continuous stability of the communication link means that within three consecutive communication detection cycles corresponding to synchronization markers, the communication relay role can receive status data packets sent by the collaborative control unit and the detection role, and there is no interruption of more than one synchronization marker between any two adjacent status data packets. The communication relay role remains stationary between adjacent occupied gates, without entering the target alarm area or occupying the evacuation path.

[0108] When the same robot dog simultaneously meets the conditions for multiple roles, the collaborative control terminal assigns roles in the order of detection, communication relay, and containment. If the target alert area is connected by only one shaded branch road, at least one robot dog is reserved as a containment role, and not all robot dogs are simultaneously assigned to the detection role. If the communication relay role cannot meet the condition of continuous communication link stability, the collaborative control terminal shortens the entry distance of the detection role, or adjusts the containment role closer to the path turning point to the communication relay role.

[0109] The collaborative control unit allocates reverse time slots for each occupied gate according to the channel occupancy cost, from highest to lowest. This reverse time slot reservation means: first, setting passage time for the gate closest to the target alarm area, which is the most difficult to evacuate after a blockage, and then setting passage time for the outer gates sequentially along the evacuation direction. Each reserved time slot at least covers the number of landing cycles expected to be required for the corresponding gate to pass through, and an additional synchronization marker is reserved as a release interval.

[0110] The detection role prioritizes acquiring the reserved time slots of occupied gates at the front end of the target alert area. The control role is only allowed to enter the vicinity of the boundary node after the detection role has passed through the corresponding occupied gate and the exit synchronization marker has appeared. The communication relay role remains between adjacent occupied gates, adjusting its position only when the communication link is interrupted or the evacuation path changes. For occupied gates with high channel occupancy costs, only one robot dog is allowed to enter within the same reserved time slot; for occupied gates with medium channel occupancy costs, only robot dogs traveling in the same direction are allowed to enter sequentially; for occupied gates with low channel occupancy costs, continuous passage is allowed as long as the safety interval is met.

[0111] The collaborative control terminal writes the role of each robot dog, the reserved time slot for entering the occupied gate, the exit synchronization flag, and the stopping position into the corresponding path, generating a staggered response path. Each robot dog moves sequentially according to the staggered response path, and uploads real-time pose information and synchronization flags when passing through the entrance and exit of the occupied gate.

[0112] When the actual release time of a vacated gate is later than the exit synchronization mark, the collaborative control terminal marks the vacated gate as delayed release and suspends the entry of its downstream robot dogs. Subsequently, the collaborative control terminal postpones the unexecuted reserved time slots according to the original arrangement. If the delayed release continues for more than 2 synchronization marks, the corrected alarm shadow map is reread; when there are other passable paths, an alternative off-peak response path is generated for the waiting robot dogs; when there are no alternative paths, the waiting robot dogs are made to stop at a position that does not obstruct the entrance of the vacated gate.

[0113] It is understandable that the detection, control, and communication relay roles can form a hierarchical response relationship around the target alarm area. Reverse time slot reservation allows occupied doors that are close to the target alarm area and difficult to evacuate to be released first, avoiding multiple robot dogs from gathering in narrow passages and ensuring the continuity of subsequent alarm detection, area control, and data transmission.

[0114] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A collaborative response method for alarm situations in a security robot dog based on multimodal perception, characterized in that, include: The visible light occlusion rate, infrared temperature difference, acoustic echo delay, gas concentration change rate and pose information collected by each robot dog are obtained to form an alarm observation sequence. Based on the alarm observation sequence and the airflow direction of the roller shutter door gap, the modal inversion boundary where the response direction of different sensing modes reverses is determined, and the alarm confidence vectors on both sides of the boundary are obtained. An alarm shadow map is established based on the alarm confidence vector and the passable path. The alarm shadow map includes shadow nodes that record risk values ​​and information missingness, and occupancy gates that record the cost of channel occupancy. Based on the aforementioned alarm shadow map, the boundary node with the largest information missing degree is selected, and at least one robot dog is controlled to perform detection and crossing. The alarm shadow map is then corrected using the observation difference before and after the crossing to obtain the target alarm area. Based on the revised alarm shadow map, each robot dog is divided into detection, control and communication relay roles, and staggered response paths are generated according to the channel occupancy cost.

2. The collaborative response method for security robot dogs based on multimodal perception as described in claim 1, characterized in that, A police situation observation sequence is formed, including: Control the robot dog to approach the roller shutter door and read the pressure at the end of the foreleg, the longitudinal micro-vibration of the body and the corresponding angular velocity of the foreleg joint; The moment when the foot pressure changes to a stable bearing state, the fuselage experiences a short-term longitudinal micro-vibration, and the corresponding forefoot joint angular velocity changes from a swinging state to a convergent state is determined as the synchronization marker. The time interval between adjacent synchronization markers is defined as the landing cycle window. Various types of sensing data are associated with a unified timestamp. The sensing data and pose information within the same landing cycle window are arranged as observation records and formed into an alarm observation sequence according to the order of movement.

3. The collaborative response method for security robot dogs based on multimodal perception as described in claim 2, characterized in that, A narrow strip observation area is set along the edge contour of the roller shutter door. Contour units that only break or become covered by water mist during the current landing cycle are marked as intermittent blanking units to obtain the visible light occlusion rate. Temperature sampling bands are set in the adjacent areas on the windward and leeward sides of the door gap to obtain the infrared temperature difference. Encoded acoustic short pulses are emitted between adjacent synchronization marks, and the acoustic echo delay is obtained based on the primary reflection from the door gap edge and the secondary reflection from the ground. A lateral sweep is performed that intersects with the airflow direction of the door gap, and the gas concentration change rate is obtained based on the gas sensor recovery segment.

4. The collaborative response method for security robot dogs based on multimodal perception as described in claim 1, characterized in that, Determine the modal inversion boundary, including: The airflow direction at the roller shutter door gap is corrected based on the drift direction of lightweight particulate matter and the recovery section of the gas sensor. Based on the pose information, a fold sampling zone is established in the environmental map. The fold sampling zone includes a windward side branch, a leeward side branch, and a branch adjacent to the door gap connecting the two. Pair the corresponding observation records in the windward and leeward branches; The stable switching position of the arrival order of primary reflection at the edge of the door gap and secondary reflection from the ground is determined as the gap-penetrating reference interval, and continuous observation windows in which the response directions of at least two sensing modes are reversed are marked as candidate inverted segments.

5. The collaborative response method for security robot dogs based on multimodal perception as described in claim 1, characterized in that, The alert confidence vectors on both sides of the boundary are obtained, including: Select a verification robot dog located within the safe passage area and control the verification robot dog to perform lateral reciprocating motion along the verification path that intersects with the airflow direction of the roller shutter door gap; According to the synchronization mark, the outward observation record and the corresponding spatial position observation record in the return are interleaved and spliced, and the candidate inverted segments that only appear in one direction are eliminated. The modal inversion boundary is extracted from the candidate inverted segments that maintain the reverse of the response direction in both the round trip. The observation records on both sides of the modal inversion boundary are collected separately, and the response direction, duration and confidence level are written in a fixed order of visible light, infrared, acoustic and gas sensing to form an alarm confidence vector.

6. The collaborative response method for security robot dogs based on multimodal perception as described in claim 1, characterized in that, Create a crime shadow map, including: Project the alarm confidence vectors on both sides of the modal inversion boundary onto the passable path, and determine the direction corresponding to the perception item with high or medium confidence level as the initial alarm direction; Perform reverse backtracking along the passable path in the opposite direction of each initial alarm; When the reverse tracing path encounters a corner, roller shutter door rail, vehicle side, vehicle bottom, or fire-fighting facility base, and an observation interruption or directional shift occurs, insert a shaded node; Shadow nodes located at the same shading structure entrance or the same path turning point are merged into composite shadow nodes, and shadow nodes directly connected to the modal inverted boundary are marked as boundary nodes.

7. The collaborative response method for security robot dogs based on multimodal perception as described in claim 6, characterized in that, Creating a crime shadow map also includes: Determine the upstream and downstream shadow nodes of each shadow node according to the reverse backtracking direction, and record the perception items that cannot be further verified by the corresponding perception mode after the path is turned as downstream missing items. Based on the pose information corresponding to the synchronization marker, the landing cycle of the missing item that is continuously crossed is converted into a continuous unchecked segment, and the information missing degree is written. Risk values ​​are written based on the retention status of upstream high-confidence sensing items before and after path reversal and the consistency of cross-modal alert directions; Path segments that cannot accommodate two robot dogs passing in parallel or safely meeting are marked as occupied gates, and the channel occupancy cost is written according to the occupancy status, the expected entry and exit order, and whether the evacuation path is cut off after the blockage.

8. The collaborative response method for security robot dogs based on multimodal perception as described in claim 1, characterized in that, Performing a probe crossing includes: Read the shadow nodes downstream along the shadow branches corresponding to each boundary node, and write the highest information missing level among the reachable downstream shadow nodes into the corresponding boundary node. When the information missing degree of multiple boundary nodes is the same, the detection crossing entrance is determined sequentially based on the number of landing cycles of consecutive crossing of the missing item, the number of downstream high-risk shadow nodes, and the channel occupation cost of the occupied gate. Control at least one detection robot dog to enter the shadow branch from the detection passage entrance, and set stopping points before the occlusion turning point, when passing through the occlusion turning point, and after leaving the occlusion turning point; Lateral sweeps and multimodal sampling were performed at each stop point.

9. The collaborative response method for security robot dogs based on multimodal perception as described in claim 8, characterized in that, Correct the police incident shading map and obtain the target police incident area, including: Using the occlusion and reversal position as a reference, the observation records downstream, in the middle and upstream are read in sequence to form a reverse staggered observation record sequence; Extraction of intermittent blanking unit position retraction, infrared temperature difference direction switching, acoustic reflection arrival sequence reset and gas recovery segment extension; When at least two observation differences appear within the same correction window covering no more than two adjacent landing cycles, a correction node is inserted, and valid observations that repeat downstream of the correction node are updated to the verified status. Update the information missingness based on the remaining missing items, update the risk value based on the consistency of cross-modal alerts, and classify the shadow nodes that continuously maintain a high risk value into the target alert area.

10. The collaborative response method for security robot dogs based on multimodal perception as described in claim 9, characterized in that, Generate off-peak response paths, including: Extract high-risk shadow nodes, boundary nodes, and occupied gates that are connected to the target alarm area from the revised alarm shadow map, and write the exit synchronization flag of each occupied gate in sequence along the evacuation direction; Robot dogs that can enter the target alarm area and have enough remaining power to meet the evacuation requirements are classified as detection roles; robot dogs that can reach the boundary node and do not occupy the evacuation path are classified as control roles; and robot dogs that can maintain two-way communication with both the collaborative control terminal and the detection role are classified as communication relay roles. Reverse time slot reservations are executed in descending order of channel occupancy cost; When the gate is delayed in being released, the reserved time slots that have not yet been executed are extended, and the staggered response path is updated according to the revised alarm shadow map.

Images