TWI505707B - Abnormal object detecting method and electric device using the same - Google Patents

Description

Abnormal object detection method and electronic device

The invention relates to an abnormal object detecting method, in particular to an abnormal object detecting method for video data and to an electronic device using the same.

The intelligent monitoring system is a monitoring system used to replace human resources to monitor the content of video data with a large amount of data. In one of its applications, the intelligent monitoring system needs to determine whether an abnormality has occurred in a scene. For example, if there is a thief in a parking lot who wants to steal the vehicle, or if someone in the station wants to grab it, etc. In general, intelligent surveillance systems need to track objects (such as people or vehicles) in these video materials. However, if these objects overlap, it will be difficult to successfully track these objects. Moreover, even if these objects are tracked, how to use these tracked trajectories to determine whether an object is an abnormal object is also a problem. Therefore, how to propose an effective abnormal object detection method is a subject of interest to those skilled in the art.

Embodiments of the present invention provide an abnormal object detecting method and using the same The electronic device of the method can accurately track a plurality of foreground objects and determine which foreground objects are abnormal objects.

An embodiment of the invention provides an abnormal object detecting method suitable for an electronic device. The abnormal object detecting method includes: acquiring a video data and separating a plurality of foreground objects from the video data; tracking the foreground objects to generate a plurality of tracks; determining, according to the tracks, a first picture in the video data, first Whether the foreground object and another foreground object overlap; if an overlap occurs, calculating a first predicted position of the first foreground object according to a Kalman filter; performing an average displacement algorithm on the first predicted position to obtain the first detect Measuring the position; correcting the first detection position according to the Kalman filter to obtain a first correction position, and updating the trajectory corresponding to the first foreground object according to the first correction position; and determining whether the foreground object is based on the trajectory Anomalous object.

In an embodiment, the step of performing an average displacement algorithm on the first predicted position to obtain the first detected position comprises: obtaining a foreground image of the first image; and, starting from the first predicted position, in the foreground image An average displacement algorithm is performed on the top to obtain the first detection position.

In an embodiment, the step of tracking the foreground object to generate a plurality of trajectories is performed according to a speckle tracking algorithm. The abnormal object detecting method further includes: if there is no overlap phenomenon, tracking the foreground object in the first picture according to the speckle tracking algorithm.

In an embodiment, the step of determining whether an overlap phenomenon occurs comprises: obtaining a new foreground object in the first picture; pairing the foreground object with a new foreground object; if at least two of the foreground objects are paired with the new foreground object, it is determined that at least two of the foreground objects overlap.

In an embodiment, the abnormal object detecting method further includes: determining, according to the trajectory, whether a second foreground object has a disappearing phenomenon in the first picture; and if a disappearing phenomenon occurs, calculating according to a Kalman filter a speed of the second foreground object, wherein a second picture is after the first picture, the second picture is different from the first picture by n pictures, and n is a positive integer; and the second is calculated according to the multiplication of the positive integer n and the velocity The foreground object is at a second predicted position of the second picture; and if the second foreground object is detected at the second predicted position, the trajectory corresponding to the second foreground object is updated according to the second predicted position.

In an embodiment, the abnormal object detecting method further includes: obtaining a training database; and performing a self-organizing incremental neural network on the training database to obtain a training model. The above step of determining whether the foreground object is an abnormal object based on the trajectory is performed according to the training model.

In an embodiment, the training model includes a plurality of nodes, and the trajectory corresponding to a third foreground object includes a plurality of trajectory information vectors. The step of determining whether the third foreground object is an abnormal object comprises: obtaining a first node of the node closest to each track information vector; obtaining a track distance between each track information vector and the corresponding first node; a farthest neighbor of each first node, and obtaining a neighbor distance between each node and the corresponding farthest neighbor; accumulating the track distances to generate an accumulated distance, and accumulating the neighbor distances to generate an accumulated neighbor distance ; and, compare this accumulated distance with this accumulated neighbor distance To determine whether the third foreground object is an abnormal object.

In an embodiment, the step of comparing the accumulated distance with the accumulated neighbor distance to determine whether the third foreground object is an abnormal object comprises: dividing the subtraction of the accumulated distance from the accumulated neighbor distance by the sum of the accumulated distance and the accumulated neighbor distance to generate a normalized distance; whether the normalized distance is greater than a first critical value to calculate an abnormal number of the third foreground object; determining whether the abnormal number is greater than a second critical value; and, if the abnormal number is greater than the second critical value, determining The third foreground object is an abnormal object.

In another aspect, an embodiment of the invention provides an electronic device including a memory and a processor. The memory stores multiple instructions. The processor is configured to execute the instructions to perform a plurality of steps: acquiring a video data and separating a plurality of foreground objects from the video data; tracking the foreground objects to generate a plurality of tracks; and determining a first picture of the video data according to the tracks Whether the first foreground object overlaps with other foreground objects; if an overlap occurs, calculating a first predicted position of the first foreground object according to a Kalman filter; performing an average displacement algorithm on the first predicted position Obtaining a first detection position; correcting the first detection position according to the Kalman filter to obtain a first correction position, and updating a trajectory corresponding to the first foreground object according to the first correction position; and determining, according to the trajectory Whether the foreground object is an abnormal object.

Based on the above, the abnormal object detecting method and the electronic device using the method provided by the embodiment of the present invention can accurately track the foreground object by combining the Kalman filter and the average displacement algorithm.

The above described features and advantages of the invention will be apparent from the following description.

110‧‧‧Electronic devices

120‧‧‧Video Information

130‧‧‧Results

140‧‧‧ processor

150‧‧‧ memory

210, 220, 230, 240, 410, 420, 430, 440, 510a, 510b, 520a, 520b, 810 ‧ ‧ images

242, 244, 324, 332, 342, 344, 352, 522, 524‧ ‧ foreground objects

Schematic diagram of 310, 320, 330, 340, 350‧‧

312, 322, 346, 354, 356, 702 ‧ ‧ new foreground objects

610, 620‧‧ Track

704‧‧‧Listing

Steps S706, S708, S710, S712, S714, S1102, S1104, S1106, S1108, S1110, S1112, S1114, S1116‧‧

716‧‧‧ current image

718‧‧‧ foreground mask

820‧‧‧ nodes

910, 920‧‧‧ icon

1001~1010‧‧‧ nodes

1031~1033‧‧‧Track information vector

1041~1043‧‧‧Track distance

1 is a block diagram of an electronic device in accordance with an embodiment.

2 is a schematic diagram showing the acquisition of a foreground object, according to an embodiment.

FIG. 3 is a schematic diagram of determining whether a foreground object overlaps according to an embodiment.

4 is a schematic diagram showing the execution of an average displacement algorithm, in accordance with an embodiment.

FIG. 5 is a schematic diagram showing a process overlap phenomenon according to an embodiment.

FIG. 6 is a schematic diagram showing a smooth trajectory according to an embodiment.

7 is a flow chart showing tracking of foreground objects in accordance with an embodiment.

FIG. 8 is a schematic diagram showing a training trajectory information vector according to an embodiment.

9 is a schematic diagram showing a training model in accordance with an embodiment.

10 is a schematic diagram showing nodes in a trajectory information vector and a training model, according to an embodiment.

FIG. 11 is a flow chart showing a method for detecting an abnormal object according to an embodiment.

1 is a block diagram of an electronic device according to an embodiment of the invention.

Referring to FIG. 1 , the electronic device 110 is configured to receive the video data 120 and determine whether one or more objects in the video data 120 are abnormal objects to generate a determination result 130 . The electronic device 110 can be implemented as a computer, a server, a distributed system, a smart phone, a tablet, any form of embedded system, or an electronic device, and the present invention is not limited to such implementations.

The electronic device 110 includes at least a processor 140 and a memory 150. A plurality of instructions are stored in the memory 150, and the processor 140 is configured to execute the instructions to generate the determination result 130. The processor 140 can be, for example, a central processing unit (CPU), a microprocessor (Microprocessor), or a digital signal processor (DSP). The memory 150 can be, for example, a random access memory or a flash memory.

First, the processor 140 receives the video material 120 and separates a plurality of foreground objects from the video material 120. For example, the processor 140 may use a Gaussian mixture model (GMM) to create a background image in the video material 120, and obtain a foreground image of each image in the video material according to the background image. The processor 140 can also remove shadow portions from the foreground image. However, the processor 140 may employ any algorithm to obtain foreground images or remove shadow portions, and the present invention is not limited to such processing algorithms. As shown in FIG. 2, the image 210 is a background image, the image 220 is a current image, the image 230 is a foreground image, and the image 240 is a foreground image after the shadow portion is removed. After the image 240 is acquired, the processor 140 retrieves the foreground objects 242 and 244 in the image 240 based on a connected component labeling algorithm. Processor 140 The identification code of the foreground objects 242 and 244, the bounding box, the size, the variable for indicating whether it is a moving object, and the variable for indicating whether or not an overlap phenomenon occurs are recorded.

Next, the processor 140 tracks all foreground objects in the video material 120 to generate a plurality of tracks. For example, processor 140 tracks these foreground objects according to a blob tracking algorithm. More specifically, all foreground objects that have been tracked are recorded in a list of BlobLists. After the processor 140 receives a new image in the video material 120, one or more new foreground objects in the new image are detected and represented as a variable BlobsNew. The processor 140 will pair these new foreground objects with the foreground objects in the list BlobList. For example, if a bounding frame of a new foreground object overlaps with a bounding box of a foreground object, it indicates that the new foreground system is paired with the foreground object. Alternatively, if the image content of a new foreground object is similar to the image content of a foreground object, it also represents that the new foreground system is paired with the foreground object, and the present invention is not limited to such pairing methods.

FIG. 3 is a schematic diagram illustrating determining whether a foreground object overlaps, according to an embodiment.

Referring to FIG. 3, in diagram 310, processor 140 detects a new foreground object 312, and the new foreground object 312 does not form a pair with the foreground object in the tandem BlobList. At this point, processor 140 will add the new foreground object 312 to the list BlobList.

In diagram 320, processor 140 detects a new foreground object 322, Only one foreground image 324 in the list BlobList is assigned. At this time, the processor 140 updates the trajectory corresponding to the foreground object 324 according to the new foreground object 324.

In diagram 330, foreground object 332 is detected by processor 140 and is not assigned to any of the new foreground objects. At this point, the processor 140 removes the trajectory corresponding to the foreground object 332 from the tandem BlobList. Moreover, the processor 140 determines that the foreground object 332 has disappeared (ie, disappeared into the new image). The following will explain how to deal with the disappearance.

In diagram 340, processor 140 detects two foreground objects 342 and 344 that are tied to new foreground object 346. At this time, the processor 140 determines that the foreground objects 342 and 344 overlap. The following will explain in detail how to handle this overlap.

In diagram 350, processor 140 detects a foreground object 352 that is paired with two new foreground objects 354, 356. At this time, the processor 140 determines that the new foreground objects 354 and 356 overlap.

On the other hand, if a foreground object is continuously tracked in n pictures (ie, no overlap or disappearance occurs), the foreground object is marked as a moving object. Where n is a positive integer, but the invention does not limit the value of n.

The following explains how to handle overlap. It is assumed that in a first picture, a first foreground object (e.g., foreground object 342, 344, 354, or 356) overlaps with other foreground objects. The processor 140 records the image content of the first foreground object, and calculates a first predicted position of the first foreground object based on the recorded trajectory and a Kalman filter. Next, The processor 140 performs a mean-shift algorithm on the first predicted position to obtain a first detected position. More specifically, the processor 140 may obtain a foreground image of the first image (eg, using a Gaussian mixture model). The processor 140 performs an average displacement algorithm on the foreground image according to the image content of the first foreground object and starting from the first predicted position to obtain the first detection position. For example, as shown in FIG. 4, the image 410 is an image obtained by the processor 140. Image 420 is the result of performing an average displacement algorithm on image 410. The image 420 contains a lot of noise due to interference from the background image. Image 430 is the foreground image of image 410 and image 440 is the result of performing an average displacement algorithm on image 430. It can be seen that in the image 440, since the background portion has been removed, the position of the foreground object can be detected more accurately.

After obtaining the first detection position, the processor 140 also corrects the first detection position according to the Kalman filter to generate a first correction position. The processor 140 updates the trajectory corresponding to the first foreground object according to the first modified position. As shown in FIG. 5, the images 510a and 510b are images before the overlap phenomenon, and the images 520a and 520b are images after the overlap phenomenon (for example, the image 520a is the first image described above). It can be seen from FIG. 5 that since the processor 140 incorporates a Kalman filter and an average displacement algorithm, the foreground objects 522 and 524 can still be tracked separately after the occurrence of the overlap phenomenon.

It is worth noting that the Kalman filter is assumed to predict the state of an object on a linear basis. The average displacement algorithm uses a kernel to calculate a value and gradually moves the core to find a Extreme position. Each pixel of image 440 is the value calculated by the core at each location on image 430. However, those skilled in the art can understand the calculation process of the Kalman filter and the average displacement algorithm, and will not be described here.

The following explains how to handle the disappearance. The processor 140 determines whether a foreground object has disappeared based on the trajectory that has been tracked (as shown in diagram 330). The reason for this disappearance may be that the foreground object moves too fast, so the processor 140 has no way to successfully track the foreground object. If a second foreground object in the first picture has disappeared, the processor 140 calculates the velocity of the second foreground object according to the Kalman filter, and uses the speed to predict the second foreground object. The position in several pictures. Specifically, the processor 140 obtains the position of the second foreground object in the t-1th picture, and predicts the position of the tth picture (ie, the first picture) according to the Kalman filter, where t is positive Integer. The position of the second foreground object at the position of the tth picture minus the position of the t-1th picture will be equal to the above speed. When the processor 140 is to predict the position of the second foreground object on the t+nth picture (also referred to as the second picture), the above speed is multiplied by a positive integer n and the second foreground object is added to the tth picture. Position, thereby obtaining a second predicted position of the second foreground object on the t+nth picture. If the processor 140 detects the second foreground object (ie, the foreground object detected at the second predicted position) in the second predicted position of the t+nth picture, it may be configured in the above-mentioned serial BlobList. The second foreground object, the processor 140 updates the trajectory corresponding to the second foreground object according to the second predicted position. In this embodiment, the positive integer n will be less than or equal to 5, but the present invention does not limit the value of the positive integer n.

As described above, the processor 140 processes the overlap and disappearance to correctly update the multiple tracks. In this exemplary embodiment, processor 140 will again use Kalman filters to smooth the trajectories. As shown in FIG. 6, trajectory 610 is the trajectory before smoothing, and trajectory 620 is the smoothed trajectory.

7 is a flow chart showing tracking of foreground objects, in accordance with an embodiment.

Referring to FIG. 7, in step S706, the processor 140 pairs the recorded foreground objects in the series 704 with the new foreground objects 702. In step S708, the processor 140 determines whether an overlap phenomenon has occurred. If the result of the determination in step S708 is no, the processor 140 tracks the foreground object according to the speckle-based tracking algorithm. If the decision result in the step S708 is YES, the processor 140 executes the Kalman filter, and performs an average displacement algorithm based on the result of the execution (step S714). Moreover, in step S714, the processor 140 filters the current image 716 according to the foreground mask 718 (eg, the result of performing the Gaussian mixture model) to obtain the foreground image, and performs an average displacement algorithm on the foreground image. The final processor 140 smoothes the trajectories according to the Kalman filter, regardless of whether or not an overlap occurs, thereby updating the series 704. The steps in Fig. 7 have been described in detail above, and will not be described again here.

Next, the processor 140 determines whether the foreground object is an abnormal object based on the smoothed trajectory. Specifically, the processor 140 first obtains a training database, which may include one or more video materials, each of which includes one or more foreground objects. The foreground objects in this training database belong to normal foreground objects. The processor 140 obtains the trajectories of all foreground objects in the training database according to the method of acquiring the trajectory described above. The processor 140 will put each of the training databases The trajectory is divided into a plurality of training trajectory information vectors including an x coordinate, a y coordinate, an x-direction velocity, a y-direction velocity, or a combination thereof of a foreground object at different times (pictures). In other words, if a foreground object appears in n pictures, the processor 140 obtains n training track information vectors. In this exemplary embodiment, the dimension of one training track information vector is 4, but the invention is not limited thereto. As shown in FIG. 8, the information vector for each training trajectory can be projected onto a node (eg, node 820) on image 810.

The processor 140 also normalizes each training trajectory information vector. In this embodiment, processor 140 divides the x coordinate by the width of the picture and divides the y coordinate by the height of the picture. The processor 140 also calculates the standard deviation of the velocity magnitudes of all of the training trajectory information vectors, and divides the velocity magnitude by the standard deviation multiplied by a constant, which can be expressed as equations (1) and (2) below.

Dx is the x-direction velocity of a training trajectory information vector, dy is the y-direction velocity, S is the velocity magnitude, and σ is the standard deviation of the velocity magnitude. c is a constant, which is a value between 2.5 and 3.5 in this embodiment, but the invention is not limited thereto. S _norm is the speed after normalization.

In addition, each speed direction is also normalized according to the following equation (3).

d _norm =( d + π )/2 π .........(3)

d is the speed direction, and d _norm is the normalized speed direction.

The processor 140 generates a normalized x-direction velocity and a y-direction velocity according to the normalized velocity magnitude and velocity direction.

The processor 140 performs a self organizing incremental neural networks (SOINN) on the training database to obtain a training model. Specifically, the processor 140 treats all the normalized training track information vectors as input values of the self-organizing incremental neural network, and the self-organizing incremental neural network removes some of the noises, and These training track information vectors are automatically classified into one or more groups, and finally a training model is generated. The training model will include multiple nodes, and each node corresponds to a training trajectory information vector. Each node may include one or more edges, and two nodes connected by one edge belong to the same group. As shown in FIG. 9, if only the x coordinate and the y coordinate are considered, the training model can be projected into a two-dimensional representation 910. If only the x-direction velocity and the y-direction velocity are considered, the training model can be projected into a two-dimensional representation 920.

Next, the processor 140 determines whether the foreground object is an abnormal object according to the training model. Specifically, if the processor 140 determines whether a third foreground object is an abnormal object, the processor 140 obtains a trajectory of the third foreground object and obtains a plurality of trajectory information vectors in the trajectory. In this embodiment, the dimension of each trajectory information vector is 4, which includes the x-coordinate, y-coordinate, x-direction velocity, and y-direction velocity of the third foreground object at a certain time (picture). Therefore, if the third foreground object appears in n pictures in the video material, the processor 140 obtains n track information vectors.

10 is a diagram showing a trajectory information vector and a section in a training model according to an embodiment. A schematic diagram of the point.

Referring to FIG. 10, it is assumed that the training model includes at least nodes 1001 to 1010, and the trajectory of the third foreground object includes the trajectory information vectors 1031 to 1033. The processor 140 retrieves the node (also referred to as the first node) closest to the trajectory information vectors 1031 to 1033 from the nodes of the training model. In this embodiment, the processor 140 determines the distance between the node and the trajectory information vector based on the Euclidean distance. For example, track information vector 1031 is closest to node 1004, track information vector 1032 is closest to node 1004, and track information vector 1033 is closest to node 1009. It is worth noting that the nodes 1001~1010 and the track information vectors 1031~1033 are four-dimensional vectors. For convenience of explanation, the two-dimensional figure is used to illustrate the Euler distance. In addition, the processor 140 obtains the trajectory distance between each trajectory information vector 1031~1033 and the corresponding node (ie, the trajectory distances 1041~1043, which are also calculated by the 尤拉 distance).

Processor 140 also retrieves one of the farthest neighbors of node 1004. Here, the farthest neighbor represents the node that is connected to node 1004 but farthest from the node. For example, node 1003 is the farthest neighbor of node 1004. Processor 140 also takes the neighbor distance between node 1003 and node 1004 (i.e., the Euler distance between node 1003 and node 1004). Similarly, processor 140 will also take the farthest neighbor of node 1009 (ie, node 1010) and take the neighbor distance between node 1009 and node 1010.

The processor 140 will accumulate these track distances 1041~1043 to generate an accumulated distance. The processor 140 also accumulates the above three neighbor distances (corresponding nodes) The neighbor distance of 1004 will be added twice to generate a cumulative neighbor distance. The processor 140 compares the accumulated distance with the accumulated neighbor distance to determine whether the third foreground object is an abnormal object.

For example, the processor 140 divides the subtraction value of the accumulated distance from the accumulated neighbor distance by the added value of the accumulated distance and the accumulated neighbor distance to generate a normalized distance, expressed as the following equation (4).

R _d =( D _sum - T _sum )/( D _sum + T _sum ).........(4)

R _d is the normalized distance, D _sum is the accumulated distance, and T _sum is the accumulated neighbor distance. The processor 140 calculates an abnormal number of times of the third foreground object according to whether the normalized distance R _d is greater than a first threshold, and determines whether the number of abnormal times is greater than a second threshold. If the number of abnormalities is greater than the second critical value, the processor 140 determines that the third foreground object is an abnormal object. These steps can be expressed by the following equations (5) and (6).

C is the number of abnormalities, R _T is the first critical value, C _T is the second critical value, and variable AO is used to indicate whether the third foreground object is an abnormal object. In this embodiment, the first critical value R _T is between 0.6 and 0.8, but the invention is not limited thereto.

FIG. 11 is a flow chart showing a method for detecting an abnormal object according to an embodiment.

Referring to FIG. 11, in step S1102, a video data is acquired and a plurality of foreground objects are separated from the video data. In step S1104, these foreground objects are tracked to generate a plurality of trajectories. In step S1106, it is determined according to the trajectory whether or not the first foreground object overlaps with other foreground objects in the first picture.

If the result of the determination in step S1106 is YES, then in step S1108, the predicted position of the first foreground object is calculated based on the Kalman filter. In step S1110, an average displacement algorithm is performed on the predicted position to obtain a detected position. In step S1112, the detected position is corrected according to the Kalman filter to obtain a corrected position, and the trajectory corresponding to the first foreground object is updated according to the corrected position.

If the result of the determination in step S1106 is NO, then in step S1114, the foreground object in the first screen is tracked according to the speckle-type tracking algorithm.

In step S1116, it is judged based on these trajectories whether or not the foreground object is an abnormal object.

The steps in Fig. 11 have been described in detail above, and will not be described again here. It should be noted that the various steps in FIG. 11 may be implemented as a code or a circuit, and the present invention is not limited thereto. Further, the flow of FIG. 11 may be implemented alone or in combination with the above embodiments, and the present invention is not limited thereto.

In summary, the abnormal object detecting method and the electronic device using the method provided by the embodiment of the present invention can utilize the Kalman filter and the average displacement algorithm to process the overlapping phenomenon to accurately track each foreground object. Moreover, through the self-organizing incremental neural network, it is possible to accurately determine whether a foreground object is an abnormal object.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention, and any one of ordinary skill in the art can make some changes and refinements without departing from the spirit and scope of the present invention. The scope of the invention is defined by the scope of the appended claims.

S1102, S1104, S1106, S1108, S1110, S1112, S1114, S1116‧‧ steps

Claims (16)

An abnormal object detecting method is applicable to an electronic device, wherein the abnormal object detecting method comprises: acquiring a video data and distinguishing a plurality of foreground objects from the video data; and tracking the foreground objects to generate a plurality of tracks; The trajectory is used to determine whether an overlap between a first foreground object and one of the other foreground objects in a first picture of the video material occurs; if the overlap occurs, Calculating a first predicted position of the first foreground object according to a Kalman filter; performing an average displacement algorithm on the first predicted position to obtain a first detected position; and correcting according to the Kalman filter The first detecting position is used to obtain a first correcting position, and the trajectory corresponding to the first foreground object among the trajectories is updated according to the first modified position; and the foreground objects are determined according to the trajectories Whether it is an abnormal object. The method for detecting an abnormal object according to claim 1, wherein the step of performing the average displacement algorithm on the first predicted position to obtain the first detected position comprises: obtaining a foreground image of the first image And performing the average displacement algorithm on the foreground image with the first predicted position as a starting point to obtain the first detected position. The method for detecting an abnormal object according to claim 1, wherein the step of tracking the foreground objects to generate a plurality of trajectories is performed according to a speckle tracking algorithm, and the abnormal object detecting method further includes: If the overlap does not occur, the foreground objects in the first picture are tracked according to the speckle tracking algorithm. The method for detecting an abnormal object according to claim 1, wherein the step of determining whether the overlapping phenomenon occurs comprises: obtaining a single new foreground object in the first picture; and performing the foreground object with the new foreground object Pairing; if at least two of the foreground objects are matched to the single new foreground object, determining that at least two of the foreground objects overlap. The method for detecting an abnormal object according to claim 1, further comprising: determining, according to the trajectories, whether a second foreground object of the foreground objects has disappeared in the first picture; And the disappearing phenomenon, the speed of the second foreground object is calculated according to the Kalman filter, wherein a second picture of the video data is after the first picture, and the second picture is different from the first picture. a picture, and n is a positive integer; multiplying the positive integer n by the speed to calculate a second predicted position of the second foreground object in the second picture; and if the second predicted position is detected at the second predicted position The second foreground object updates the trajectory corresponding to the second foreground object among the trajectories according to the second predicted position. The method for detecting an abnormal object according to claim 1, further comprising: obtaining a training database; performing a self-organizing incremental neural network on the training database to obtain a training model, wherein The trajectory to determine whether the foreground objects are abnormal objects is performed according to the training model. The abnormal object detecting method according to claim 6, wherein the training model includes a plurality of nodes, the foreground objects include a third foreground object, and the trajectory corresponding to the third foreground object includes a plurality of a trajectory information vector, wherein the step of determining whether the third foreground object is an abnormal object comprises: obtaining a first node of the nodes closest to each of the trajectory information vectors; obtaining each of the trajectory information vectors and corresponding a track distance between the first nodes; obtaining a farthest neighbor of each of the first nodes, and obtaining a neighbor distance between each of the nodes and the corresponding farthest neighbor; accumulating the tracks The distance is generated to generate an accumulated distance, and the neighbor distances are accumulated to generate an accumulated neighbor distance; and the accumulated distance is compared with the accumulated neighbor distance to determine whether the third foreground object is an abnormal object. Such as the method for detecting an abnormal object described in claim 7 of the patent scope, wherein The step of determining the third foreground object as an abnormal object by comparing the accumulated distance with the accumulated neighbor distance includes: subtracting the subtraction distance from the accumulated distance from the accumulated neighbor distance by dividing the accumulated distance and the accumulated neighbor distance Adding a value to generate a normalized distance; calculating an abnormal number of times of the third foreground object according to whether the normalized distance is greater than a first critical value; determining whether the abnormal number of times is greater than a second critical value; and if the abnormal number of times If the second threshold is greater than the second threshold, the third foreground object is determined to be an abnormal object. An electronic device comprising: a memory for storing a plurality of instructions; and a processor for executing the instructions to perform a plurality of steps: obtaining a video data and distinguishing from the video data Tracking the foreground objects to generate a plurality of trajectories; determining, according to the trajectories, a first foreground object of the foreground objects in a first picture of the video material, and other foreground objects Whether one of the overlapping phenomena occurs; if the overlapping phenomenon occurs, calculating a first predicted position of the first foreground object according to a Kalman filter; performing an average displacement algorithm on the first predicted position Obtaining a first detection position; correcting the first detection position according to the Kalman filter to obtain a first correction Positioning, and updating the trajectory corresponding to the first foreground object among the trajectories according to the first modified position; and determining whether the foreground objects are abnormal objects according to the trajectories. The electronic device of claim 9, wherein the step of performing the average displacement algorithm on the first predicted position to obtain the first detected position comprises: obtaining a foreground image of the first image; The first predicted position is a starting point to perform the average displacement algorithm on the foreground image to obtain the first detected position. The electronic device of claim 9, wherein the step of tracking the foreground objects to generate a plurality of trajectories is performed according to a speckle tracking algorithm, the steps further comprising: if the overlapping phenomenon does not occur And tracking the foreground objects in the first picture according to the speckle tracking algorithm. The electronic device of claim 9, wherein the step of determining whether the overlapping phenomenon occurs comprises: acquiring a new foreground object in the first picture; pairing the foreground objects with the new foreground object; And at least two of the foreground objects are associated with the new foreground object, and then the at least two of the foreground objects are determined to have the overlapping phenomenon. The electronic device of claim 9, wherein the steps further comprise: Determining, according to the trajectories, whether a second foreground object of the foreground objects has a disappearing phenomenon in the first picture; if the disappearing phenomenon occurs, calculating a second foreground object according to the Kalman filter Speed, wherein a second picture of the video material is after the first picture, the second picture is different from the first picture by n pictures, and n is a positive integer; the positive integer n is multiplied by the speed, Calculating a second predicted position of the second foreground object in the second picture; and if the second foreground object is detected at the second predicted position, updating the corresponding one of the tracks according to the second predicted position The trajectory of the second foreground object. The electronic device of claim 9, wherein the steps further comprise: obtaining a training database; performing a self-organizing incremental neural network on the training database to obtain a training model, wherein The step of determining whether the foreground objects are abnormal objects is performed according to the training model. The electronic device of claim 14, wherein the training model comprises a plurality of nodes, the foreground objects comprise a third foreground object, and the trajectory corresponding to the third foreground object comprises a plurality of trajectory information vectors, The step of determining whether the third foreground object is an abnormal object comprises: obtaining a first node of the nodes that is closest to each of the track information vectors; Obtaining a trajectory distance between each of the trajectory information vectors and the corresponding first node; obtaining a farthest neighbor of each of the first nodes, and obtaining each of the nodes and the corresponding farthest neighbor a neighbor distance between the two; accumulating the track distances to generate an accumulated distance, and accumulating the neighbor distances to generate an accumulated neighbor distance; and comparing the accumulated distance with the accumulated neighbor distance to determine whether the third foreground object is Anomalous object. The electronic device of claim 15, wherein comparing the accumulated distance with the accumulated neighbor distance to determine whether the third foreground object is an abnormal object comprises: adding the accumulated distance to the accumulated neighbor distance Offset, divided by the added value of the accumulated distance and the accumulated neighbor distance to generate a normalized distance; and based on whether the normalized distance is greater than a first critical value, an abnormal number of times of the third foreground object is calculated; Determining whether the abnormal number is greater than a second threshold; and if the abnormal number is greater than the second threshold, determining that the third foreground object is an abnormal object.