Method for detecting evidence and integrated system for detecting evidence using the method

Abstract

This invention provides a method for detecting evidence left behind by a person and an integrated system for detecting such evidence. [Solution] The object detection method includes step A, which is a preprocessing step used for acquiring site images and foreground detection; step B, which is an object pairing and tracking step that detects objects in the original RGB images and selects objects that frequently appear in the monitored environment; and step C, which is an object detection step that detects whether an object is a common object at the site based on the object's overlap rate, displacement, and dwell time. This enables monitoring of whether objects at a site in a public space are artifacts.

Description

[Technical Field]

[0001] This invention relates to a detection method for monitoring whether an object in a public space is a relic, and to an integrated system for detecting relics using this method. [Background technology]

[0002] Previous studies have used object detection methods to define eight categories, such as furniture, mattresses, and sofas. However, in real-world scenarios, defining objects with eight categories is insufficient because of the diversity of the objects themselves, making object feature detection methods applicable only to specific situations. Another study showed that if the color of an object is similar to or indistinct from the background, it can lead to a low signal-to-noise ratio or detection errors. For example, if a discarded item is a dark-colored object, it may appear similar in color to the asphalt road surface after imaging, increasing the difficulty of recognition. The two studies mentioned above demonstrate that similarity between the appearance color and background color of objects makes recognition difficult in terms of object detection.

[0003] Furthermore, the environment for detecting abandoned objects is affected by weather and reflection. In terms of weather, if it rains during monitoring, the noise generated by the rain increases the complexity of recognition. Another scenario to consider is an environment with water, glass, or mirrors. For example, in a rainy scene, if the rider who dumped the abandoned object is wearing a raincoat, and there is water at the monitoring site after the rain, images generated by puddles on the ground, the water surface at the site, glass, and reflective mirrors, while not debris, can hinder recognition and increase the difficulty of recognition. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2025-60131 [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] The objective of the present invention is to improve upon the problems of the prior art described above by providing a more novel "method for detecting abandoned objects" and an "integrated system for detecting abandoned objects using the method," which can detect the act of leaving abandoned objects in public spaces and can be extended to industrial production sites to monitor whether there are abandoned objects such as tools, buckets, personal belongings, or people (who have fallen at the scene due to an accident) at inspection points. For example, it can be applied to detecting illegal dumping or disposal of garbage by residents in neighborhoods or alleyways. In the detection scenario, the detection of abandoned objects outdoors is selected because outdoor scenes have more variations in light and shadow than indoor scenes, and abandoned objects also have diverse characteristics depending on the scenario. After implementation of this system, it can be extended in the future to monitor production sites and indoor public spaces. [Means for solving the problem]

[0006] The present invention, "Method for Detecting Remaining Objects," is realized through the following implementation details, and its technical features are as follows. Step A is a preprocessing step used for acquiring field images and detecting foregrounds, and includes: (1) acquiring a pixel and initializing a background model that includes a set of values ​​previously acquired at the same location or nearby; (2) acquiring a new pixel and comparing it with the values ​​in the background model, and if a sufficiently large number of values ​​are close to the value of the new pixel, the new pixel is considered to belong to the background, and otherwise, the new pixel is considered to belong to the foreground; (3) if the pixel belongs to the background, updating a specific value in the background model as the new pixel value with a predetermined probability and propagating the new pixel value to the background model of adjacent pixels with a predetermined probability; (4) repeating (2) and (3) until all pixels have been processed; and (5) detecting field objects; Step B is an object pairing and tracking step that detects objects in the original RGB image and selects objects that frequently appear in the monitoring environment; and Step C is a remnant detection step that detects whether an object is a common object at the site based on the overlap rate, displacement, and dwell time of the object.

[0007] In another preferred embodiment, the conditions for detecting whether or not an object is a trace object in the trace detection step are: 1) the object does not move continuously, and one movement threshold (thr moved ) defines an object residence threshold (thr stay-up-duration ) defines that the object must remain at the site for a certain period of time, and if its existence time exceeds a threshold, it is considered a remnant; and 3) it is not an identifiable object in the scene, and one object detection function f(x) object‐detection It defines that the input is a tracked object, and whether the output is true or false means whether a general object was found, and if either does not meet the criteria for a relic, the existence duration stay_up ga thr stay‐up‐durationIt is mandatory to check if the value is greater than / 2, and if so, to reset the movement amount, which means that the object is continuing to move, thus resolving situations in which the object continues to move, such as in scenarios where the object is dumped from a vehicle or thrown a certain distance from a distance. In the residual object detection process, the conditions for detecting whether an object is a residual object include: 1) the object does not move continuously, a movement threshold (thrmoved) is defined, and if the object falls below the movement threshold, it may be a residual object. 2) the object remains in the same place. A threshold is defined for the duration an object stays. To be considered a residual object, the object must remain at the scene for a longer time than the threshold. 3) for objects that cannot be recognized in the scene, an object detection function f(x)object-detection is defined, which takes the object to be tracked as input and outputs a boolean value indicating whether a common object was detected. If either does not meet the criteria for a residual object, it is checked whether the existence time durationstay_up is greater than thrstay-up-duration / 2. If it is greater, the movement amount, which indicates that the object is still moving, needs to be reset to resolve situations where the residual object moves continuously, such as when the residual object is thrown from a car or thrown from a distance. [Effects of the Invention]

[0008] The advantages and innovations of the method for detecting remains in public spaces developed using the present invention are as follows. 1) It provides an architecture for improving foreground detection algorithms, captures dynamic movements in videos and combines them with object detection mechanisms, emphasizes the foreground through methods of resolution reduction, noise reduction, and expansion, and solves the problem of object detection errors in changing environments. 2) It uses an algorithm to pair and track objects, solves the problem of misidentifying remains due to light and shadow reflections, and uses this to trace back and find the discarders. 3) It uses an artificial intelligence training output model to predict and extract the features of tracking images, and detects remains using the cosine similarity comparison method, thus solving the problem that it is difficult to recognize remains using object detection-based methods.

Brief Description of the Drawings

[0009] <0OO0094>It is the main structural diagram of the integrated system of the present invention. [Figure 2] It is the main flowchart of the method for detecting remains of the present invention. [Figure 3] It is the main flowchart of the preprocessing step of the present invention. [Figure 4] It is the main flowchart of the object pairing and tracking step of the present invention. [Figure 5] It is the main flowchart of the step for detecting remains of the present invention. [Figure 6] It is the main flowchart of the system integration of the present invention.

Modes for Carrying Out the Invention

[0010] The description of the content of the present invention above and the description of the following implementation methods are used to demonstrate and explain the principles and methods of the present invention, and to further explain the scope of the patent application of the present invention.

[0011] Embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Following standard industry practice, various functions are not depicted to scale and are used for illustrative purposes only. In fact, the dimensions of components can be arbitrarily enlarged or reduced to clearly illustrate the features of the embodiments of the present invention.

[0012] As shown in Figure 1, the object detection integrated system 1 described in the present invention comprises, in order, a surveillance camera 10, a communication base station 11, a network exchange 12, a cloud computing server 13, and a mail server 14 connected via a network. The cloud computing server 13 further includes at least one artificial intelligence engine (AI Engine) 130, a stream server 131, a database server 132, and an event center 133. The surveillance camera 10 is used to acquire real-time surveillance video at a fixed point and transmit the video data via the communication network. The video is transmitted to the cloud computing server 13 via the communication base station 11 and the network exchange 12, and the stream image is received by the stream server 131. After an object is detected through the object detection mechanism and method of the present invention contained within the artificial intelligence engine 130, the data is transmitted in real time to the event center 133, event processing is triggered, stored in the notification mechanisms of the database server 132 and the mail server 14, and notified to a personal computer or mobile phone used by the observer. The artificial intelligence engine 130 has a file reading mechanism and provides offline analysis capabilities to address the scenarios of existing surveillance systems at the site.

[0013] As shown in Figure 2, the method for detecting artifacts according to the present invention mainly comprises step A, which is a preprocessing step (SA); step B, which is an object pairing and tracking step (SB); and step C, which is an artifact detection step (SC). Of these, the preprocessing step (SA) is responsible for detecting and enhancing the foreground, the object pairing and tracking step (SB) analyzes objects in preceding and succeeding frame images using an algorithm, and the artifact detection step (SC) detects artifacts in the scene based on object overlap rate, object size difference, similarity, displacement, dwell time, and whether the object is unusual in the scene. As described above, after detecting artifacts, the data is transmitted in real time to the event center 133 to initiate event processing, is stored in the notification mechanisms of the database server 132 and mail server 14, notified to monitors using a PC or mobile phone, and system integration (SD) is completed.

[0014] The preprocessing (SA) step, as shown in Figure 3, further includes multiple mechanisms such as image reading, foreground detection, resolution reduction, image noise reduction and foreground enhancement, object detection, filtering, and masking.

[0015] In the image reading portion, since the present invention can be applied to both offline analysis and online real-time analysis, the image reading portion can be divided into reading offline files for offline analysis and reading real-time images from the surveillance camera 10, but is not limited to this. After reading, foreground detection and object detection steps are performed.

[0016] The foreground detection section includes the following steps and processes. (1) Obtain a pixel and initialize a background model that includes a set of values ​​previously obtained at the same location or in the vicinity. (2) A new pixel is obtained and compared with the value of the background model. If a sufficient number of values ​​are close to the value of the new pixel, the new pixel is considered to belong to the background; otherwise, the new pixel is considered to belong to the foreground. (3) If a pixel belongs to the background, a specific value in the background model is updated as a new pixel value with a predetermined probability, and the new pixel value is propagated to the background model of adjacent pixels with a predetermined probability. (4) Repeat steps (2) and (3) until all pixels have been processed.

[0017] Furthermore, considering the screen input from the camera or offline video file, in the initial stages of the system, no background is established. At this time, the first image is the background after resolution reduction and image noise reduction. After the second image is input, the algorithm is executed on both the second and first images. Using the results of the above image subtraction, the changed areas in the image can be obtained. This is the result of foreground detection. Areas with small remaining differences are the background, while areas with changes may be potential foreground elements. Foreground detection can be achieved in this way.

[0018] After foreground detection is complete, a mask is generated, and the content marked by the mask represents the foreground and background after subtraction of the two images. The foreground is marked in white, and the background is marked in black. BGS_Frame represents the result after performing background subtraction (hereinafter referred to as Background Subtraction, BGS) on a single frame.

[0019] After foreground detection, the BGS_Frame contains a considerable number of tiny white dots, and these tiny dots need to be filtered using several processing techniques. Reducing the resolution in preprocessing removes relatively small white dots, thereby reducing the subsequent computational load.

[0020] Image noise reduction is a preprocessing step, and its function is to reduce the influence of the external environment on the image, but the operation method is different from reducing the resolution. In this invention, the method of image noise reduction calculates the number of white pixels in the connected region of BGS_Frame and sets one noise reduction judgment threshold thr denoise Set the number of white pixels in the connected area, WPixel BGS‐CC ga thr denoiseIf it is smaller than, it is regarded as noise, and all white areas (foreground, value 1) regarded as noise are changed to black (background, value 0). By this method, noise is filtered, and the filtering process is as shown in Equation 1 below. If WPixel BGS-CC < thr denoise , set WPixel BGS-CC = 0 (Equation 1)

[0021] After noise reduction, perform dilation twice to strengthen the foreground range to obtain the foreground mask BGS_Mask. When detecting objects later, it will be possible to more accurately distinguish between the foreground and the background.

[0022] Object detection, filtering, and masking are methods for capturing people or vehicles that initiate residual actions within the system. The specific detection process includes the following object detection, filtering, and masking processes.

[0023] <Object Detection> This object detection detects objects within the RGB original image. The object detection mechanism of the present invention quotes the YOLO official model, selects objects that frequently appear in the monitoring environment, and detects, for example, people, cars, bicycles, motorcycles, dogs, cats, etc.

[0024] <Filtering> After detection, a series of objects within the image are obtained. At this time, it is necessary to emphasize the detected target objects. The emphasizing method is to first initialize a completely black image without any mask represented as total_mask, and for each detected object, use YOLO to mark the bounding box (BBox) of the object. The area enclosed by the BBox is the foreground (FG) and the background (BG). The sum of the two represents the size of the BBox, and WPixel YOLONext, the ViBe algorithm is used to perform foreground-background segmentation (BGS). After the calculation, a mask obj_mask of the foreground object is generated. The region size of the mask obj_mask is calculated and WPixel is used. BGS Obtain the foreground WPixel. BGS The size of BBox is WPixel YOLO After obtaining the foreground WPixel BGS BBox whole WPixel YOLO The proportion is calculated. WPixelBGS It is expressed as follows. The calculation process is as shown in Equation 2, and if the percentage is too low, it is not a monitoring object and therefore needs to be filtered. The filtering threshold is Thr WPixelRatio It is expressed as: Ratio WPixelBGS Th WPixelRatio If the value is less than 0, filtering is necessary. The filtering method is to set the entire range of the object to 0, as shown in Equation 3. TIFF2026121362000002.tif1783 (Formula 2) If Ratio WPixelBGS < Thr WPixelRatio , set WPixel BGS = 0 (Equation 3)

[0025] <Masking Process> The masking process is divided into three steps: retaining detected objects, retaining foreground objects, and noise reduction. After filtering, all remaining objects, obj_mask, are the objects that need to be continuously detected. As shown below, these are sequentially ORed with total_mask to retain all objects that need to be detected. for each obj_Mask : total_mask = total_mask ∨ obj_mask

[0026] After processing the detected objects, the foreground detection results are processed again, and as a result, total_mask and BGS_mask are ORed as shown below. total_mask = total_mask ∨ BGS_mask

[0027] After processing, a final noise reduction process is performed. The purpose of image noise reduction is to eliminate some small dots to remove noise. The purpose of the noise reduction process here is to enhance the foreground. The specific processing method involves four steps: blurring, binarization, dilation, and filtering. The purpose of the first three steps, blurring, binarization, and dilation, is to reduce noise and detail, remove unwanted details or objects, and enhance and make edges or features in the image stand out.

[0028] Finally, in the filtering section, all objects remaining after processing up to this point are objects that need to be tracked later, so a threshold thr is used to filter the objects. drop Set the number of pixels in the area covered by the mask of any object. obj_mask ga thr drop If it is less than 0, the object's area is set to 0 and removed from total_mask, as shown in Equation 4. If WPixel obj_mask < thr drop , set WPixel obj_mask = 0 (Equation 4)

[0029] If the system does not detect any objects during object detection, total_mask will perform noise reduction processing on the foreground BGS_mask as follows: total_mask = denoise(BGS_mask)

[0030] After the object detection and processing steps described above, the object list BBox and total_mask are output and used in subsequent object tracking calculations.

[0031] The object pairing and tracking (SB) process involves continuously updating the objects that need to be tracked, and the process is shown in Figure 4. The list of objects to be continuously tracked is represented by tracking_list. tracking_list is a list used to store the objects currently being tracked. tracking_list is updated in every frame. If an object is considered new or does not match any of the objects currently being tracked, that object is added to tracking_list. At the same time, if an object has not matched for several consecutive frames, that object is removed from tracking_list.

[0032] The following defines the variables x, y, and z used in the tracking process. x: This is one object within the tracking list and is used to pair with object y in the interest box. y: One object within the interest_box (the object box list of the current frame). z: This is an object in the tracking list that is used to check whether it is covered by other objects or whether it does not match the foreground in multiple consecutive frames.

[0033] After the process is executed, the preprocessing steps continue to output BBox and total_mask, perform interconnected region analysis on the objects, and generate a list of one interesting object, interest_box. interest_box is a list used to store all objects in the current frame.

[0034] The `interest_box` is obtained through connected region analysis. Each object is a rectangular box indicating its position within the image. Connected regions are implemented by using the OpenCV `connectedComponentsWithStats()` function to analyze and obtain information such as the object's boundary rectangle and area.

[0035] After the analysis of the communication region is complete, the algorithm is run for each y to check if there is a corresponding x that can be paired. In the object pairing section, interest_box is the input object list, and tracking_list is the list of objects to track. When both are paired and the system is initialized, if there are no objects to track, all objects in interest_box are added to tracking_list. When a second image is input, pairing begins. The variables used in object pairing in this invention are the intersection area (IoU), size difference (size-diff), and similarity. The intersection area (IoU) refers to the size of the intersection area of ​​the two paired objects. This variable is between 0 and 1, where 0 means there is no intersection between the two objects, and 1 means the two objects overlap. The closer the IoU value is to 1, the more the two objects match. The size difference refers to the difference in area between the two paired object BBoxes, and the area is in pixels. The size-diff is calculated by subtracting the pixel counts of the two objects from each other, taking the absolute values, and excluding those with relatively large differences. The resulting size-diff value is between 0 and 1. The closer the value is to 0, the more the two objects are identical. The calculation method is as shown in Equation 5. TIFF2026121362000003.tif18110 (Formula 5)

[0036] Regarding the similarity aspect, this invention references the concept of person re-identification, specifically the Torchreid library provided by OpenCV. This library has the functionality to train models through deep learning, and the trained models have the ability to extract features from images.

[0037] After feature extraction from the objects in `interest_box` and `tracking_list`, feature vectors are generated. The similarity between the two feature vectors is obtained by performing a cosine similarity operation. The similarity is a numerical value between -1 and 1, where 1 represents the most similar and -1 represents the least similar. This is used for object pairing.

[0038] After pairing, if pairing is successful, the system continues to determine if it is a relic. If pairing fails, it means it is a new object. It is added to the tracking_list for continued tracking in the next frame. After excluding the paired object, the system checks if the object is blocked. If the object box overlaps with the extent of other object boxes in the current frame, it means the object is blocked and needs to be checked if it is a relic. If it is not overlapped, the system checks if the similarity is greater than 0.9. If it is greater than 0.9, the object is currently being tracked and its duration is checked. stay_upThis means that the similarity is +1. If the similarity is less than 0.9, it means that the object in the display screen has changed and the similarity is not high, and if an object does not match the foreground for several consecutive frames, it means that the object is in the background and should be deleted. After the pairing process, the paired objects in interest_box and tracking_list are updated, the objects in interest_box replace the corresponding objects in tracking_list based on the pairing relationship, and finally, the object box can be marked and output to the artifact detection process.

[0039] Figure 5 shows the flow of the abandoned object detection process. The data input to the abandoned object detection process is a list of paired objects in interest_box. In this detection process, each paired object is first compared one by one with interest_box and tracking_list to check whether the intersection area (IoU), size difference (size-diff), and similarity match as the same object. This solves the problem of false positives of highly homogeneous objects. For example, if multiple pieces of garbage in garbage bags of the same color are discarded at the same time, a misjudgment may occur when discarding the garbage. After detection, if any of the intersection area (IoU), size difference (size-diff), or similarity (similarity) does not meet the set threshold conditions, the object may not be the target object and needs to be separated from the original image, made into a new object, and added to tracking_list. This process is very important. When a person moves away from an abandoned object, the two objects on the image begin to separate, and this is done by continuously tracking the unpaired object and adding the separated object. Conversely, if all three thresholds are passed, it means that the object is currently being tracked. After updating the box information in the object box, further detection is needed to determine if it is a remnant.

[0040] To detect whether the current object is a relic, it is first necessary to clearly define what a relic is. Based on the nature of the relic itself, the detection conditions are set as follows: 1) The object does not move continuously, and there is a movement threshold (thr moved ) defines an object residence threshold (thr stay-up-duration ) defines that the object must remain at the site for a certain period of time, and if the existence time is greater than the threshold, it is considered a remnant. This is as shown in Equation 7. 3) It is not an identifiable object in the scene (e.g., a person, cat, dog, vehicle, etc.), and one object detection function f(x) object‐detection This defines a function where the input is a tracked object, and the output being true or false indicates whether a general object was detected, as shown in Equation 8. This prevents people, cats, and dogs from remaining at the scene of a crime scene for a certain period of time and being mistakenly identified as crime objects. If motion obj < thr moved , set Flag trash-c1 = true (Formula 6) If value stay_up > thr stay-up-duration , set Flag trash-c2 = true (Formula 7) If f(object) object-detection = false , set Flag trash-c3 = true (Formula 8)

[0041] If any of the above formulas 6 to 8 does not meet the standard for the remains, then the duration of existence stay_up ga thr stay‐up‐durationIt checks if it is greater than / 2. The movement amount needs to be reset to indicate that the object is still moving. The purpose of this design is to resolve situations where an object is dumped from a vehicle or when a remnant is dumped a certain distance from afar. After the remnant detection mechanism has finished, the tracking_list is sent to the system and integrated to determine if there are any events.

[0042] The system integration part is shown in Figure 6. After the artifact detection process is completed, the event center 133 enters the decision process. The event center 133 processes each individual object in the tracking list depending on whether it is marked as an artifact. If it is an artifact, the event information is processed, the event is saved, and a notification is sent.

[0043] Based on the above, the advantages and innovations obtained by using the public space object detection system and method developed in this invention are as follows: 1) Provides an architecture that improves the foreground detection algorithm, capturing dynamic motion in video and combining it with the object detection mechanism, enhancing the foreground by methods of resolution reduction, noise reduction, and enhancement, and solving the problem of object detection errors in situations where the environment changes. 2) Uses the algorithm to pair and track objects, solving the problem of misidentification of objects due to light and shadow reflections, and using this to trace back and find the person who dumped the object. 3) Solves the problem of the difficulty of object detection-based methods in recognizing objects by predicting and extracting features from tracked images using an artificial intelligence trained output model and detecting objects using the cosine similarity comparison method. [Explanation of Symbols]

[0044] SA Pre-treatment Process SB Object Pairing and Tracking Process SC (Science Control) process for detecting evidence left behind. SD System Integration 1. Integrated system for detecting evidence left behind 10 surveillance cameras 11. Communication base stations 12 Network Switches 13 Cloud Computing Servers 14 Mail Server 130 Artificial Intelligence Engines 131 Stream Servers 132 Database Servers 133 Event Center

Claims

1. This is a preprocessing step used for acquiring on-site images and detecting the foreground. (1) Obtain a pixel and initialize a background model that includes a set of values ​​previously obtained at the same position or nearby, (2) A new pixel is obtained and compared with the value of the background model. If the value of the new pixel is close to the value of the background, the new pixel is considered to belong to the background; otherwise, the new pixel is considered to belong to the foreground. (3) If a pixel belongs to the background, a specific value in the background model is updated as a new pixel value with a predetermined probability, and the new pixel value is propagated to the background model of adjacent pixels with a predetermined probability. (4) Repeat steps (2) and (3) until all pixels have been processed, (5) Detecting on-site objects, Step A includes, Step B is an object pairing and tracking process that detects objects in the original RGB image and selects objects that frequently appear in the monitoring environment. Step C is a debris detection process that detects whether the object is a common object at the site based on the overlap rate, displacement, and residence time of the aforementioned object. A method for detecting evidence, including evidence.

2. After the detection of the foreground is completed, a mask is generated, and the contents marked by the mask are the foreground and background after subtracting the two images from each other, with the foreground marked in white and the background marked in black, and BGS_Frame represents the result of one frame after performing Background Subtraction (BGS), and BGS_Frame contains fine white dots, and by reducing the resolution, the amount of subsequent computation is reduced by removing relatively small fine white dots in the preprocessing step. The method for detecting traced objects according to claim 1.

3. The number of white pixels in the communication region of the BGS_Frame is calculated, and the threshold for noise reduction judgment is determined. ｄｅｎｏｉｓｅ Set the number of white pixels in the communication area WPixel ＢＧＳ‐ＣＣ gathr ｄｅｎｏｉｓｅ The method for detecting traced objects according to claim 2, wherein if the value is less than a certain value, it is considered noise, and all white areas (foreground, numerical value 1) considered to be noise are changed to black (background, numerical value 0) to filter the noise.

4. The object pairing and tracking process continuously updates the objects that need to be tracked, and the list of objects to be continuously tracked is represented by `tracking_list`, which is a list for storing the currently tracked objects. The method for detecting a lost object according to any one of claims 1 to 3, wherein in each frame, the tracking_list is updated, and if an object is considered new or does not match an object currently being tracked, the object is added to the tracking_list, and at the same time, if an object has not been paired in multiple consecutive frames, the object is removed from the tracking_list.

5. After the object pairing and tracking process is completed, BBox and total_mask are output. The object is analyzed in a connected region, and a list of objects of interest, called interest_box, is generated. This interest_box is a list for storing all objects in the current frame. The object detection method according to claim 4, wherein interest_box is obtained by communication region analysis, and each object is a single rectangular box indicating the position of the object in the image.

6. The variables used for pairing the aforementioned objects include three parameters: intersection area (IoU), size difference (size-diff), and similarity. The intersection area (IoU) refers to the size of the intersection area of ​​the two paired objects. This variable ranges from 0 to 1, where 0 indicates no intersection between the two objects, and 1 indicates overlap. The closer the IoU value is to 1, the more the two objects match. The size difference refers to the difference in the area of ​​the BBoxes of two paired objects. Area is measured in pixels. The size difference is calculated by subtracting the pixel counts of the two paired objects from each other, then taking the absolute value. Excluding values ​​that are relatively large between the two values, the resulting size difference is between 0 and 1. A value closer to 0 indicates a closer match between the two paired objects. The method for detecting traced objects according to claim 5, wherein the similarity is obtained by generating feature vectors by extracting features between an object in interest_box and an object in tracking_list, performing a cosine similarity operation on the two feature vectors, and the similarity is a numerical value between -1 and 1, where 1 represents the most similar and -1 represents the least similar.

7. After the object has been paired, if the pairing is successful, it is determined whether it is a relic; if the pairing fails, it means it is a new object and is added to the tracking_list to continue tracking in the next frame; after removing the paired object, it is checked whether the object is blocked; if the object box overlaps with the extent of other object boxes in the current frame, it means the object is blocked and it is necessary to check whether it is a relic; if they do not overlap, it is checked whether the similarity is greater than 0.9; if it is greater than 0.9, the object is currently being tracked and its duration is checked. stay_up ) means +1, and if the similarity is less than 0.9, it means that the object in the display screen has changed and the similarity is not high, and if an object does not match the foreground for several consecutive frames, it means that the object is in the background and should be deleted. The method for detecting traced objects according to claim 6, wherein after the pairing process, the paired objects in interest_box and tracking_list are updated, the objects in interest_box replace the corresponding objects in tracking_list based on the pairing relationship, and finally, the object box is marked and output to the traced object detection process.

8. The data input to the aforementioned object detection process is a list of objects in interest_box that have completed pairing. For each paired object, interest_box and tracking_list are compared one by one to check whether the intersection area (IoU), size difference (size-diff), and similarity correspond to the same object. After detection, if any of the three conditions—IoU, size difference (size-diff), and similarity—do not meet the set threshold conditions, it means that the object may not be the target object, and it needs to be separated from the original image, made into a new object, added to tracking_list, and tracking should continue when the next image is input. Conversely, if all three thresholds are passed, it means that the object is currently being tracked, and the method for detecting a lost item according to claim 7 further detects whether it is a lost item after updating the box information of the object box.

9. The conditions for detecting whether or not something is a personal object in the aforementioned object detection process are: 1) The object does not move continuously, and there is one movement threshold (thr moved This defines that objects below the movement threshold can be considered relics, 2) The duration of an object's stay in the same place, and the threshold for the duration of an object's stay (thr stay-up-duration ) defines that the object must remain at the site for a certain period of time, and if its existence time exceeds a threshold, it is considered a remnant. 3) It is not an identifiable object in the scene, and one object detection function f(x) ｏｂｊｅｃｔ‐ｄｅｔｅｃｔｉｏｎ This defines that the input is a tracked object, and whether the output is true or false means whether a general object was found. Includes, If either does not meet the criteria of the relic, the existence duration ｓｔａｙ_ｕｐ is checked to see if it is greater than thr ｓｔａｙ‐ｕｐ‐ｄｕｒａｔｉｏｎ / 2, and if it is, it is essential to reset the amount of movement, which means the object is continuing to move, to solve the situation where the relic continues to move, such as the scenario of discarding the relic from the vehicle or throwing the relic a certain distance from afar. The relic detection method according to claim 8

10. An integrated system for detecting evidence by any one of claims 1 to 9, comprising surveillance cameras, a communication base station, a network exchange, a cloud computing server, and a mail server, wherein the cloud computing server comprises at least one artificial intelligence engine (AI Engine), a stream server, a database server, and an event center, the surveillance cameras are used to acquire real-time surveillance video at a fixed point and transmit the video data over a communication network, the video is transmitted to the cloud computing server via the communication base station and network exchange, the stream images are received by the stream server, and after detecting evidence using an evidence detection mechanism and method contained within the AI ​​Engine, the data is transmitted in real time to the event center, event processing is initiated, stored in the notification mechanism of the database server and mail server, and notified to a personal computer or mobile phone used by the observer.

Images