A visual-based garbage station personnel off-duty detection method

Abstract

This invention relates to a vision-based method for detecting absenteeism of garbage station personnel. This method detects garbage station staff by monitoring video streams, solving the problem of low efficiency associated with traditional manual supervision. The technical solution includes: acquiring video streams and identifying staff wearing armbands; using the DeepSORT algorithm for personnel tracking; simultaneously detecting vehicle targets; calculating the minimum distance between a person and a vehicle when the person disappears, and determining whether they have left their post based on the distance; calculating a dwell time score when the person and vehicle are close, determining whether they are seeking shelter from the cold; tracking associated vehicles, and determining the absence status based on whether the person reappears and their positional relationship; when an armband is missing, determining whether armband transfer has occurred by calculating the armband-limb positional relationship score, movement displacement score, and relative armband displacement score. This invention can accurately identify absenteeism of garbage station personnel, improving supervision efficiency.

Description

Technical Field

[0001] This invention relates to visual inspection methods, specifically to a vision-based method for detecting the absence of personnel at waste disposal stations. Background Technology

[0002] With the acceleration of urbanization and the increasing awareness of environmental protection, the standardized management of garbage stations, as an important component of urban environmental management, is receiving increasing attention. The on-duty status of garbage station staff directly affects the normal operation of the station and environmental sanitation; therefore, effective monitoring of staff absence from their posts is of great significance. Currently, vision-based personnel absence detection technology has been widely applied in various fields. By capturing images of the monitored area through cameras and combining them with computer vision and artificial intelligence algorithms, automatic identification and early warning of personnel absence from their posts can be achieved.

[0003] In existing technologies, CN115512298A discloses a method for identifying personnel leaving their posts in a centralized control area based on a target tracking algorithm within a defined region. This method uses Yolov5 to detect targets within a monitored image and then uses the DeepSORT tracking algorithm to track the movement of the targets, thereby determining whether personnel have left their posts. It can track continuously for 24 hours, ensuring the continuity of the departure detection. CN118397656A proposes a method and system for detecting personnel leaving their posts based on surveillance video. It collects a personnel target detection dataset, trains a target detection model, and uses the trained target detection model to detect personnel targets within a defined area in real time, obtaining the position coordinates and confidence scores of personnel targets within the area. By analyzing the detection results of consecutive detection frames, it reduces the impact of missed personnel in a single frame on the final alarm result.

[0004] CN114519532A discloses a machine vision-based method and related device for personnel status detection. This method combines deep neural networks, multi-target tracking algorithms, identity recognition algorithms, and foreground region recognition, enabling the tracking and identification of multiple personnel in a monitored video frame, thereby achieving the detection of multiple people leaving their posts / sleeping on duty within the monitored video area. CN117372953A proposes a method for leaving posts based on a secondary target tracking algorithm, combining two tracking algorithms to perform two target tracking operations, thus improving accuracy. CN116486339A discloses a method and device for identifying suspicious personnel in a data center campus. By acquiring video data, it determines the location features and crowd distance features of the target personnel, generates multi-dimensional features, and inputs them into a preset suspicious personnel identification model to obtain the identification results of the target personnel.

[0005] However, existing leave-of-post detection technologies primarily focus on whether personnel are within designated areas or rely on simple target tracking to determine if personnel have left their posts. They fail to effectively distinguish between "reasonable shelter from the cold" and "genuine absence from duty" by garbage station staff. In frigid regions, garbage station staff may briefly enter vehicles to escape the cold. While this behavior temporarily removes them from their posts, it is not a true absence from duty. Traditional leave-of-post detection systems often misinterpret this shelter-from-the-cold behavior as absence from duty, leading to numerous false alarms.

[0006] Furthermore, existing technologies have limitations in personnel identification, particularly when armbands are passed between guards, making it difficult for the system to accurately track the actual guards. Current solutions typically rely on auxiliary devices such as temperature sensors to determine if personnel are sheltering from the cold inside the vehicle, but this not only increases the system's complexity and cost, but also compromises the reliability of these devices in extremely cold environments.

[0007] Therefore, there is an urgent need for a purely visual method for detecting absence from duty. This method should be able to understand the spatial semantic relationship between "personnel-vehicle-work area" and intelligently distinguish between the behavior of garbage station guards seeking shelter from the cold and their actual absence from duty, without relying on any auxiliary sensors. It should also be able to accurately handle complex scenarios such as the passing of armbands, thereby improving the accuracy and practicality of absence detection. Summary of the Invention

[0008] To address the technical challenge of traditional visual absence detection systems failing to distinguish between "reasonable shelter from the cold" and "genuine absence from duty," and to achieve accurate absence detection using a pure visual AI solution, this paper presents a vision-based method for detecting absences from garbage station personnel.

[0009] The technical solution adopted by this invention to solve its technical problem is: a vision-based method for detecting personnel leaving their posts at garbage stations, comprising the following steps: S1. Personnel Identification and Tracking: Acquire the video stream of the garbage station monitoring system and detect personnel targets in the image; determine whether the personnel are guards using the armband recognition model; continuously track the identified guards using DeepSORT and obtain the center coordinates of the personnel bounding box in each frame; S2. Vehicle Detection and Information Acquisition: Synchronously detect vehicle targets in the screen and acquire the position of the vehicle's bounding box (xc). i,j yc i,j wc i,j hc i,j ), where j is the vehicle number, (xc i,j yc i,j ) represents the coordinates of the left vertex of the box, wc i,j hc is the width of the box. i,j The height of the box; S3. Personnel Disappearance Judgment and Distance Calculation: When the tracked guard disappears in frame i1, calculate the minimum distance d1 between the person and each vehicle before disappearance; based on the relationship between the minimum distance d1 and the set threshold t1, enter different judgment branches. S4. Direct Departure Judgment: When d1 > ts1, the personnel are determined to be in a state of being off duty; if the guard is not re-identified after a set time t1, the personnel are confirmed to have left their post. S5. Vehicle Association and Dwelling Degree Analysis: When d1 At time ts1, determine the vehicle corresponding to the minimum distance and calculate the score g1 of the degree of stay of the personnel near the vehicle; based on the relationship between the degree of stay score g1 and the threshold ts2, determine whether the personnel are in a state of being off duty or suspected of being in a state of seeking shelter from the cold. S6. Tracking and re-determination of vehicles suspected of being in a state of seeking shelter from the cold: When a vehicle is suspected of being in a state of seeking shelter from the cold, features are extracted from the associated vehicles and ByteTrack is used for tracking. During vehicle tracking, based on whether the personnel reappear, their positional relationship with the vehicle, and time thresholds, it is determined whether the personnel got off the vehicle to return, left their post without entering the vehicle, or left their post due to the vehicle leaving. S7. Target switching determination when armband is missing: When no armband is detected in the tracking personnel's box and there are other personnel with armbands, determine whether to enter the target switching verification based on the relative distance between the two; by reviewing multiple frames, calculate the positional relationship score between the armband and each limb part, the motion displacement score, and the relative displacement score of the armband; determine whether armband transfer has occurred based on the score results, and decide whether to switch the tracking target.

[0010] Preferably, the armband recognition model in step S1 is trained using the YOLO target detection algorithm. By detecting the armband features in the personnel image, the model identifies the guards wearing armbands, thus distinguishing them from ordinary personnel.

[0011] Preferably, in step S1, DeepSORT is used to continuously track the identified guards and obtain the center coordinates (x, y) of the personnel bounding box in each frame i. i y i Establish personnel time-series trajectories to achieve cross-frame identity preservation for the same person.

[0012] Preferably, the formula for calculating the minimum distance d1 in step S3 is: ; in, Let these be the coordinates of the center of the personnel bounding box before the guard disappears in frame i1. is the left vertex coordinate of the box for the j-th vehicle in the i1-th frame; in step S4, when d1 > ts1, it is determined that the person has left the post, where ts1 is the set distance threshold with a default value of 30; if the guard is not recognized after time frame t1, it is confirmed that the person has left the post, where t1 is the off-duty judgment time set according to the business with a default value of 15000.

[0013] Preferably, in step S5, the calculation method of the stay degree score g1 is as follows: , where n1 is the length of the stay time frame set according to the business with a default value of 15.

[0014] Preferably, the g2 i is the stay judgment score for the i-th frame, and the calculation formula is: , When g1 < ts2, it is determined that the person has left the post, and when g1 ≥ ts2, it is determined that the person is in a suspected cold-avoiding state, where ts2 is the set stay degree threshold with a default value of 0.9.

[0015] Preferably, in step S6, feature extraction is performed on the associated vehicle and ByteTrack is used for tracking. The end conditions for vehicle tracking include the following three situations: (1) At the i2-th frame, the guard reappears in the camera, the vehicle is always in the camera, and the distance between the person and the vehicle satisfies , then it is determined that the guard gets off the vehicle and the person is tracked again; (2) At the i2-th frame, the guard reappears in the camera, the vehicle is always in the camera, and the distance between the person and the vehicle satisfies , then it is determined that the guard does not enter the vehicle; if i2 - i1 > t1, it is determined that the person has left the post, otherwise the person is tracked again; [[ID=三十]] (3) The vehicle disappears in the camera. After time frame t1, if the guard or the tracked vehicle is not recognized, it is determined that the person has left the post. If the guard is detected, the person is tracked again. If the tracked vehicle is detected, the vehicle is tracked again.

[0016] Preferably, during the vehicle tracking in step S6, the re-identification determination of the person is made by comparing the position relationship between the person and the vehicle when the person reappears, combined with the time threshold t1, to distinguish the three situations of "the person gets off the vehicle and returns to the work post", "the person leaves the post without entering the vehicle", and "the person leaves the post due to the vehicle leaving".

[0017] Preferably, the target switching determination when the armband is missing in step S7 includes the following sub-steps: Step S71: When no armband is detected in the frame of the tracked guard in the i3th frame, determine whether there is another person in the frame with an armband detected; if not, continue tracking the original target person. ; Step S72: If another person wearing an armband exists Calculate the relative distance between the two people ,like Then continue tracking the original target, if Then proceed to target switching verification, where ts3 is the set threshold for personnel proximity, with a default value of 40; Step S73: Obtain n2 consecutive frames of playback footage, where n2 is the playback time frame length set by the service, with a default value of 25; obtain the personnel bounding box (x1) of each frame. i y1 i w1 i h1 i Armband coordinates (xb) i yb i ), (slx) i sly i ), right shoulder coordinates (srx) i ,sry i Left elbow coordinate (elx) i ,ely i Right elbow coordinates (erx) i ery i Left-handed coordinates (hlx) i ,hly i Right-handed coordinates (hrx) i hry i ); Step S74: Calculate the left arm position score g3 i Right arm position score g4 i Left elbow position score g5 i Right elbow position score g6 i Left-hand position score g7 i Right-hand position score g8 i And the armband position score g9 i .

[0018] Preferably, the final determination method for target switching in step S7 is as follows: Calculate the motion displacement score: ; Calculate the armband displacement score g 11 : , When g is satisfied 10 0 and g 11 0 and At that time, the personnel responsible for determining the location pass on the armband and track down the new target guard. Conversely, if the personnel determine that the armband was not passed, the personnel tracking the original target should be identified as guards. .

[0019] Preferably, the left arm position score is g3. i Right arm position score g4 i Left elbow position score g5 i Right elbow position score g6 i Left-hand position score g7 i Right-hand position score g8 i And the armband position score g9 i The calculation formula is as follows: , , where is the projected coordinate; , , where is the projected coordinate; , , where is the projected coordinate; , , where is the projected coordinate, and ts4 is the offset threshold set, with a default value of 5; , Calculate the armband position score g9 i : .

[0020] The beneficial effects of this invention are as follows: it achieves accurate off-duty detection using a pure visual AI solution, eliminating the need for auxiliary equipment such as temperature sensors, thus reducing system costs; by understanding the spatial semantic relationship between "personnel-vehicle-work area," it can intelligently distinguish between legitimate attempts to avoid the cold and genuine off-duty behavior; it exhibits better reliability in extremely cold environments, avoiding the failure of auxiliary sensors in harsh conditions; and it achieves accurate, reliable, and low-cost on-duty monitoring. Compared to traditional visual off-duty detection systems, this invention effectively solves the problem of being unable to distinguish between "legitimate attempts to avoid the cold" and "genuine off-duty behavior," and does not require additional auxiliary equipment such as temperature sensors, reducing system complexity and cost, and improving system reliability in extremely cold environments. Detailed Implementation

[0021] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments. The content mentioned in the embodiments is not intended to limit the present invention.

[0022] Example 1 A camera is installed directly opposite the garbage collection station pole, centered on the station. The camera captures the video stream from the garbage collection station. Station staff wear uniform armbands, and the system identifies personnel in the scene in real time based on the camera's video stream.

[0023] The vision-based method for detecting absenteeism of waste station personnel includes the following steps: S1. Personnel Identification and Tracking: Acquire the video stream of the garbage station monitoring system and detect personnel targets in the image; determine whether the personnel are guards using the armband recognition model; continuously track the identified guards using DeepSORT and obtain the center coordinates of the personnel bounding box in each frame; S2. Vehicle Detection and Information Acquisition: Synchronously detect vehicle targets in the screen and acquire the position of the vehicle's bounding box (xc). i,j yc i,j wc i,j hc i,j ), where j is the vehicle number, (xc i,j yc i,j ) represents the coordinates of the left vertex of the box, wc i,j hc is the width of the box. i,j The height of the box; S3. Personnel Disappearance Judgment and Distance Calculation: When the tracked guard disappears in frame i1, calculate the minimum distance d1 between the person and each vehicle before disappearance; based on the relationship between the minimum distance d1 and the set threshold t1, enter different judgment branches. S4. Direct Departure Judgment: When d1 > ts1, the personnel are determined to be in a state of being off duty; if the guard is not re-identified after a set time t1, the personnel are confirmed to have left their post. S5. Vehicle Association and Dwelling Degree Analysis: When d1 At time ts1, determine the vehicle corresponding to the minimum distance and calculate the score g1 of the degree of stay of the personnel near the vehicle; based on the relationship between the degree of stay score g1 and the threshold ts2, determine whether the personnel are in a state of being off duty or suspected of being in a state of seeking shelter from the cold. S6. Tracking and re-determination of vehicles suspected of being in a state of seeking shelter from the cold: When a vehicle is suspected of being in a state of seeking shelter from the cold, features are extracted from the associated vehicles and ByteTrack is used for tracking. During vehicle tracking, based on whether the personnel reappear, their positional relationship with the vehicle, and time thresholds, it is determined whether the personnel got off the vehicle to return, left their post without entering the vehicle, or left their post due to the vehicle leaving. S7. Target switching determination when the armband is missing: When no armband is detected within the box of the tracked person and there are other armband-wearing persons, it is determined whether to enter the target switching verification based on the relative distance between the two; by backtracking multiple frames of images, calculate the position relationship score, action displacement score, and relative displacement score of the armband with each limb part; determine whether armband transfer has occurred based on the score results and decide whether to switch the tracking target.

[0024] The armband recognition model in step S1 is trained using the YOLO object detection algorithm. By detecting the armband features in the image of the person, the guard wearing the armband is recognized to distinguish the guard from the ordinary person.

[0025] In step S1, DeepSORT is used to continuously track the recognized guards, and the center coordinates (x i , y i ) of the person box in each frame i are obtained to establish a person time-series trajectory and achieve cross-frame identity maintenance of the same person.

[0026] The calculation formula for the minimum distance d1 in step S3 is: ; where [[ID=2\0]]is the center coordinate of the person box before the disappearance of the guard in the i1-th frame, is the left vertex coordinate of the box of the j-th vehicle in the i1-th frame; in step S4, when d1 > ts1, it is determined that the person leaves the post, where ts1 is the set distance threshold, with a default value of 30; if no guard is recognized after the time frame t1, it is confirmed that the person leaves the post, where t1 is the departure judgment time set according to the business, with a default value of 15000.

[0027] The calculation method for the stay degree score g1 in step S5 is: , where n1 is the length of the stay time frame set according to the business, with a default value of 15.

[0028] g2 i is the stay judgment score in the i-th frame, and the calculation formula is: , When g1 < ts2, it is determined that the person leaves the post, and when g1 ts2, it is determined that the person is in a suspected cold-avoiding state, where ts2 is the set stay degree threshold, with a default value of 0.9.

[0029] In step S6, feature extraction is performed on the associated vehicle and ByteTrack is used for tracking. The end conditions for vehicle tracking include the following three cases: (1) In the i2th frame, the guard reappears in the camera, the vehicle remains in the camera, and the distance between the guard and the vehicle satisfies If so, it is determined that the guards should get off the vehicle and the personnel should be tracked again; (2) In the i2th frame, the guard reappears in the camera, the vehicle remains in the camera's view, and the distance between the guard and the vehicle satisfies If i2-i1>t1, it is determined that the guard has not entered the vehicle; if i2-i1>t1, it is determined that the guard has left his post, otherwise the guard will be tracked again. (3) If the vehicle disappears from the camera, after time frame t1, if no guard or tracking vehicle is detected, it is determined that the personnel have left their posts. If the guard is detected, the personnel will be tracked again. If the tracking vehicle is detected, the vehicle will be tracked again.

[0030] In step S6, the personnel re-identification determination during vehicle tracking is carried out by comparing the positional relationship between the personnel and the vehicle when the personnel reappear, combined with the time threshold t1, to distinguish three situations: "personnel get off the vehicle and return to their work post", "personnel leave their post without entering the vehicle", and "personnel leave their post due to the vehicle leaving".

[0031] The target switching determination when the armband is missing in step S7 includes the following sub-steps: Step S71: When no armband is detected in the frame of the tracked guard in the i3th frame, determine whether there is another person in the frame with an armband detected; if not, continue tracking the original target person. ; Step S72: If another person wearing an armband exists Calculate the relative distance between the two people ,like Then continue tracking the original target, if Then proceed to target switching verification, where ts3 is the set threshold for personnel proximity, with a default value of 40; Step S73: Obtain n2 consecutive frames of playback footage, where n2 is the playback time frame length set by the service, with a default value of 25; obtain the personnel bounding box (x1) of each frame. i y1 i w1 i h1 i Armband coordinates (xb) i yb i ), (slx) i sly i ), right shoulder coordinates (srx) i ,sry i Left elbow coordinate (elx) i ,ely i Right elbow coordinates (erx) i eryi Left-handed coordinates (hlx) i ,hly i Right-handed coordinates (hrx) i hry i ); Step S74: Calculate the left arm position score g3 i Right arm position score g4 i Left elbow position score g5 i Right elbow position score g6 i Left-hand position score g7 i Right-hand position score g8 i And the armband position score g9 i .

[0032] The final determination method for target switching in step S7 is as follows: Calculate the motion displacement score: ; Calculate the armband displacement score g 11 : ; When g is satisfied 10 0 and g 11 0 and Personnel are assigned to pass armbands to those who are tracking new targets. Conversely, if the personnel determine that the armband was not passed, the personnel tracking the original target should be identified as guards. .

[0033] Left arm position score g3 i Right arm position score g4 i Left elbow position score g5 i Right elbow position score g6 i Left-hand position score g7 i Right-hand position score g8 i And the armband position score g9 i The calculation formula is as follows: , , where is the projected coordinate; , , where is the projected coordinate; , , where is the projected coordinate; , , where is the projected coordinate, and ts4 is the offset threshold set, with a default value of 5; ; Calculate the armband position score g9 i : .

[0034] Through the above steps, this method can effectively detect the off-duty status of garbage station guards and handle various complex scenarios, including personnel leaving their posts, seeking shelter from the cold, and passing on armbands, thereby improving the intelligence level and supervision efficiency of garbage station management.

[0035] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A visual-based trash station personnel off-duty detection method, characterized in that: Includes the following steps: S1. Personnel Identification and Tracking: Acquire the video stream of the garbage station monitoring system and detect personnel targets in the image; determine whether the personnel are guards using the armband recognition model; continuously track the identified guards using DeepSORT and obtain the center coordinates of the personnel bounding box in each frame; S2. Vehicle Detection and Information Acquisition: Synchronously detect vehicle targets in the screen and acquire the position of the vehicle's bounding box (xc). i,j yc i,j wc i,j hc i,j ), where j is the vehicle number, (xc i,j yc i,j ) represents the coordinates of the left vertex of the box, wc i,j hc is the width of the box. i,j The height of the box; S3. Personnel Disappearance Judgment and Distance Calculation: When the tracked guard disappears in frame i1, calculate the minimum distance d1 between the person and each vehicle before disappearance; based on the relationship between the minimum distance d1 and the set threshold t1, enter different judgment branches. S4. Direct Departure Judgment: When d1 > ts1, the personnel are determined to be in a state of being off duty; if the guard is not re-identified after a set time t1, the personnel are confirmed to have left their post. S5. Vehicle Association and Dwelling Degree Analysis: When d1 At time ts1, determine the vehicle corresponding to the minimum distance and calculate the score g1 of the degree of stay of the personnel near the vehicle; based on the relationship between the degree of stay score g1 and the threshold ts2, determine whether the personnel are in a state of being off duty or suspected of being in a state of seeking shelter from the cold. S6. Tracking and re-determination of vehicles suspected of being in a state of seeking shelter from the cold: When a vehicle is suspected of being in a state of seeking shelter from the cold, features are extracted from the associated vehicles and ByteTrack is used for tracking. During vehicle tracking, based on whether the personnel reappear, their positional relationship with the vehicle, and time thresholds, it is determined whether the personnel got off the vehicle to return, left their post without entering the vehicle, or left their post due to the vehicle leaving. S7. Target switching determination when armband is missing: When no armband is detected in the tracking personnel box and there are other personnel with armbands, determine whether to enter the target switching verification based on the relative distance between the two; by reviewing multiple frames, calculate the positional relationship score between the armband and each limb part, the motion displacement score, and the relative displacement score of the armband. The scoring results determine whether a badge transfer has occurred and whether to switch the tracking target.

2. The vision-based method for detecting personnel leaving their posts at garbage stations according to claim 1, characterized in that: The armband recognition model in step S1 is trained using the YOLO object detection algorithm. By detecting the armband features in the personnel image, it identifies the guards wearing armbands, thus distinguishing them from ordinary personnel.

3. The visual-based trash station staff off-duty detection method of claim 1, wherein: The step S1 uses DeepSORT to continuously track the identified guards, obtains personnel box center coordinates (x i , y i ) of each frame i, establishes a personnel time sequence track, and realizes cross-frame identity maintenance of the same personnel.

4. The visual-based trash station staff off-duty detection method of claim 1, wherein: The formula for calculating the minimum distance d1 in step S3 is as follows: ； in, Let these be the coordinates of the center of the personnel bounding box before the guard disappears in frame i1. Let d1 be the coordinates of the left vertex of the bounding box of the j-th vehicle in the i1-th frame; in step S4, when d1>ts1, it is determined that the personnel have left their posts, where ts1 is the set distance threshold; if no guard is identified after time frame t1, it is confirmed that the personnel have left their posts, where t1 is the time for leaving posts set according to the business.

5. The visual-based trash station staff off-duty detection method of claim 1, wherein: The method for calculating the dwell time score g1 in step S5 is as follows: ， Where n1 is the dwell time frame length set according to the service, with a default value of 15.

6. The vision-based method for detecting personnel leaving their posts at a garbage station according to claim 5, characterized in that: The g2 i The dwell judgment score of the i-th frame is calculated by the following formula: ， When g1 < ts2, it is determined that the person has left the post. When g1 < ts2, it is determined that the person is in a suspected cold-avoiding state, where ts2 is the set threshold of the staying degree.

7. The visual-based trash station attendant off-duty detection method of claim 1, wherein: In step S6, features of the associated vehicles are extracted and tracked using ByteTrack. The vehicle tracking ends under the following three conditions: (1) The guard reappears in the camera at the i2th frame, the vehicle is always in the camera, and the distance between the guard and the vehicle satisfies Then it is determined that the guard gets off the vehicle, and the guard is tracked again. (2) In the i2th frame, the guard reappears in the camera, the vehicle remains in the camera's view, and the distance between the guard and the vehicle satisfies If so, it is determined that the guards did not enter the vehicle; If i2-i1>t1, then the personnel are determined to have left their posts; otherwise, the personnel are tracked again. (3) If the vehicle disappears from the camera, after time frame t1, if no guard or tracking vehicle is detected, it is determined that the personnel have left their posts. If the guard is detected, the personnel will be tracked again. If the tracking vehicle is detected, the vehicle will be tracked again.

8. The visual-based trash station attendant off-duty detection method of claim 7, wherein: In step S6, the re-identification of personnel during vehicle tracking is determined by comparing the positional relationship between the personnel and the vehicle when the personnel reappear, combined with the time threshold t1, to distinguish three situations: "personnel get off the vehicle and return to their work post", "personnel leave their post without entering the vehicle", and "personnel leave their post due to the vehicle leaving".

9. The visual-based trash station attendant off-duty detection method of claim 1, wherein: The target switching determination in step S7 when the armband is missing includes the following sub-steps: Step S71, when the i3 frame is identified to the tracking of the guard box does not detect the armband, determine whether there is another personnel frame in the picture detected armband; If not, continue to track the original target personnel ; Step S72, if there is another badge person , calculate the relative distance between the two persons , if , continue to track the original target, if <ts3, enter target switching verification, wherein ts3 is a set person proximity threshold value, and the default value is 40; Step S73: Obtain playback images of n2 consecutive frames, where n2 is the playback time frame length set by the service. Get the personnel bounding box (x1) for each frame i y1 i w1 i h1 i Armband coordinates (xb) i yb i ), (slx) i sly i ), right shoulder coordinates (srx) i ,sry i Left elbow coordinate (elx) i ,ely i Right elbow coordinates (erx) i ery i Left-handed coordinates (hlx) i ,hly i Right-handed coordinates (hrx) i hry i ); Step S74: Calculate the left arm position score g3 i Right arm position score g4 i Left elbow position score g5 i Right elbow position score g6 i Left-hand position score g7 i Right-hand position score g8 i And the armband position score g9 i .

10. The visual-based trash station attendant walk-off detection method of claim 9, wherein: The final determination method for target switching in step S7 is as follows: Calculate the motion displacement score: ； Computing the pennant displacement score g 11 : ， When g is satisfied 10 0 and g 11 0 and At that time, the personnel responsible for determining the location pass on the armband and track down the new target guard. ; Conversely, the judge does not pass the armband and follows the original target guard .

Images