CN112949615A - Multi-target tracking system and method based on fusion detection technology - Google Patents
Multi-target tracking system and method based on fusion detection technology Download PDFInfo
- Publication number
- CN112949615A CN112949615A CN202110519994.8A CN202110519994A CN112949615A CN 112949615 A CN112949615 A CN 112949615A CN 202110519994 A CN202110519994 A CN 202110519994A CN 112949615 A CN112949615 A CN 112949615A
- Authority
- CN
- China
- Prior art keywords
- target
- frame image
- coordinate information
- tracking
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/48—Matching video sequences
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a multi-target tracking method fusing detection technology. The method comprises the steps of firstly modeling monitoring videos and pictures, stably and accurately outputting the category and position information of a concerned target and tracking. Specifically, firstly, a target detector is used for classifying and positioning targets in a picture, then a tracker based on motion modeling is used for continuously predicting a target track, then a detection area and a motion prediction area are input into an appearance feature acquisition network to obtain respective matching features, and data association between the detection area and the motion prediction area and correction of a final tracking result frame are completed according to the target position and the matching features. The method can accurately and quickly detect the target information, meanwhile, the motion prediction position and matching feature network fully utilizes the target motion and the appearance feature, and the result of multi-target tracking is more accurate through the correction of matching and frames.
Description
Technical Field
The invention belongs to the technical field of intelligent identification, and particularly relates to a multi-target tracking method based on a fusion detection technology.
Background
The multi-target tracking method is widely applied to engineering, and plays a very key role in the work of monitoring road traffic, illegal behavior identification and the like. Given relevant videos, the traditional tracking method needs to manually initialize a tracked target frame, and with the development of deep learning, the tracking technology based on the detection of the neural network is increasing.
The detection method used in the current detection algorithm mostly adopts a one-stage or two-stage target detection algorithm, the one-stage target detection algorithm is that the features are directly mapped to the coordinate information and the category information of the target, the two-stage target detector firstly carries out the coarse positioning of all foreground targets as a regional candidate network and then inputs the regional candidate network into a classifier for the relocation and classification. The one-stage target detection algorithm has the advantages of high detection speed but low accuracy; the two-stage target detection algorithm has the advantages of high accuracy and low detection speed.
The multi-target tracking algorithm locates a plurality of targets in a video sequence based on a detection result, and forms a track through the corresponding relation between frames, so that the multi-target tracking algorithm focuses on difference learning among different target individuals. The common multi-target tracking framework gives association to the detection result of unified identity by measuring the distance between the appearance and the motion information of the detection targets and combining the association of the previous frame and the next frame. The method proves the effectiveness to a certain extent, but the design flow means that the tracking performance depends on the detection result unilaterally, and the visual characteristic with discrimination is obtained, so that a complex mechanism and a huge calculation amount are introduced, the tracking result is limited, and the efficiency is also deficient.
Disclosure of Invention
In order to solve the problems, the invention aims to provide a multi-target tracking method based on a fusion detection technology, and the multi-target tracking method has the characteristics of high specific accuracy and high detection rate.
In order to achieve the purpose, the invention adopts the following technical scheme: a multi-target tracking system based on fusion detection technology, the system comprising: the system comprises a deep convolutional neural network, a target detection network, a Kalman filter model, a matching feature acquisition network and a tracking result correction network; the output end of the deep convolutional neural network is connected with the input end of a target detection network, the output end of the target detection network is respectively connected with the input end of a matched feature acquisition network and the input end of a Kalman filter model, the output end of the Kalman filter model is connected with the input end of the matched feature acquisition network, and the output end of the matched feature acquisition network is connected with the input end of a tracking result correction network.
The invention also provides a tracking method of the multi-target tracking system based on the fusion detection technology, which specifically comprises the following steps:
(1) collecting video frame images;
(2) outputting a first frame image and a second frame image of the video frame image to a deep convolutional neural network to obtain the characteristics of each target in the first frame image and the second frame image;
(3) respectively inputting the characteristics of each target in the first frame image and the second frame image into a target detector network, outputting the confidence score, the category type, the category score and the coordinate information of each target in the first frame image and the second frame image, calculating the product of the confidence score and the category score of each target, and reserving the category type of the target with the product higher than a threshold value and the corresponding coordinate information;
(4) inputting the category type of the reserved first frame image and the corresponding coordinate information into a Kalman filter model for target tracking prediction, and predicting the coordinate information corresponding to the category type in the second frame image;
(5) inputting the second frame image, the coordinate information corresponding to the category type in the predicted second frame image and the reserved coordinate information of the second frame image into a matching feature acquisition network to obtain a predicted shape matching feature and a shape matching feature;
(6) calculating distance measurement according to the coordinate information of the second frame image, the coordinate information of the predicted second frame image, the predicted appearance matching characteristic and the appearance matching characteristic, and matching the coordinate information of the second frame image with a target corresponding to the coordinate information of the predicted second frame image by using a Hungarian algorithm;
(7) inputting the predicted shape matching feature and the shape matching feature of the second frame image obtained in the step (5) and the matched coordinate information and predicted coordinate information into a tracking result correction network, and outputting corrected coordinate information of multi-target tracking;
(8) and (5) sequentially repeating the steps (2) to (7) on the subsequent frame image until all the video frame images are tracked.
Further, the matching feature obtaining network is composed of a backbone network and a feature mapping module, and the step (5) is specifically: inputting the second frame image into a backbone network to obtain a complete feature map, cutting the coordinate information corresponding to the category type in the predicted second frame image and the part of the reserved coordinate information of the second frame image at the corresponding position in the feature map, respectively inputting the two parts of features obtained by cutting into a feature mapping module, and outputting a 1 x 128-dimensional predicted shape matching feature and a shape matching feature.
Further, the distance metric
: the calculation process of (2) is as follows:
wherein the content of the first and second substances,
represents the secondCoordinate information P of frame image and coordinate information of predicted second frame image
The ratio of the intersection of (a) to the union thereof,
in order to predict the shape-matching features,
is a shape matching feature.
Further, if the matching fails in the step (6), continuously predicting a certain number of frames of the motion prediction result until the matching is achieved in a certain frame, or continuously predicting the certain number of frames, and then stopping the track operation without prediction, wherein the matching with the detection result cannot be achieved yet; if there is no detection result matching the motion prediction result, re-recognition is performed to determine whether to restart the old target or to newly initialize a new target.
Further, the tracking method further comprises: after the tracking task of the frame image is completed, the shape matching feature F of the tracking corresponding detection result is usedPUpdating the shape matching feature F of each tracking trackT：
Wherein the content of the first and second substances,
it is indicated that the learning rate is,
representing the shape-matching features of the trace before processing the current frame,
and showing the track shape matching characteristics after the track corresponding to the target shape information of the current frame is updated.
Compared with the prior art, the invention has the beneficial effects that: the detection method of the multi-target tracking method adopts an advanced target detection network, and provides more accurate initialization and observation values for target tracking. In the aspect of motion prediction, the invention adopts a Kalman filtering-based motion modeling method, only uses motion information of a target to perform iterative update, accurately predicts the position of an object and introduces extremely low calculation cost, and on the basis of fusion of shape matching characteristics and position information based on a detection result and a tracking result, a more accurate tracking result can be obtained through a tracking result correction network.
Drawings
FIG. 1 is a flow chart of a multi-target tracking method based on fusion detection technology.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. On the contrary, the invention is intended to cover alternatives, modifications, equivalents and alternatives which may be included within the spirit and scope of the invention as defined by the appended claims. Furthermore, in the following detailed description of the present invention, certain specific details are set forth in order to provide a better understanding of the present invention. It will be apparent to one skilled in the art that the present invention may be practiced without these specific details.
The invention provides a multi-target tracking system based on fusion detection technology, which comprises: the system comprises a deep convolutional neural network, a target detection network, a Kalman filter model, a matching feature acquisition network and a tracking result correction network; the output end of the deep convolutional neural network is connected with the input end of a target detection network, the output end of the target detection network is respectively connected with the input end of a matched feature acquisition network and the input end of a Kalman filter model, the output end of the Kalman filter model is connected with the input end of the matched feature acquisition network, and the output end of the matched feature acquisition network is connected with the input end of a tracking result correction network.
Referring to fig. 1, a flowchart of a tracking method of a multi-target tracking system based on a fusion detection technology is shown, which specifically includes the following steps:
(1) obtaining surveillance video and video image frames through a road traffic surveillance camera
，
(2) Image of video frame
The first frame image and the second frame image are output to a deep convolutional neural network
Obtaining the characteristics of each target in the first frame image and the second frame image
(ii) a ResNet-50 is adopted as the deep convolution neural network in the invention
。
(3) The characteristics of each object in the first frame image and the second frame image are compared
Separately input into a network of target detectors
In the first frame image and the second frame image, the confidence score S, the category type C and the category score of each object in the first frame image and the second frame image are output
And coordinate information P
In which is definedThe category is
= { pedestrian, mei qu takeaway electric vehicle, hungry take-out electric vehicle, non-takeaway electric vehicle },
is the number of categories. And calculating a confidence score S and a category score for each object
Will be higher than a threshold value
Class of object of (2)
And corresponding coordinate information P are reserved;
(4) setting an observation vector of a target by motion modeling position prediction based on a custom uniform velocity linear Kalman filter model
Defining a modeled target state vector for the center coordinates, width and height, respectively, of the target bounding box
The last four items are respectively the moving speed, the width and the high change rate of the central point of the target enclosing frame in the x and y directions, the category type C of the reserved first frame image and the corresponding coordinate information P are input into a Kalman filter model for target tracking prediction, and the coordinate information corresponding to the category type in the second frame image is predicted
(ii) a Specifically, the method comprises the following steps:
(4.1) category of first frame image to be retained
And corresponding coordinate information P, initializing an observation vector of the target
Setting the motion speed and the width-height change rate to be 0;
(4.2) passing the motion state of the first frame image
Predicting the second frame image state with the state transition matrix H
：
(ii) a Error matrix through first frame image
State transition matrix H and noise covariance matrix Q, error matrix for predicting second frame image
；
(5) Predicting the coordinate information corresponding to the category in the second frame image
And inputting the coordinate information P of the reserved second frame image into a matching feature acquisition network to obtain a predicted appearance matching feature
And profile matching features
(ii) a The matching feature acquisition network consists of a backbone network F' and a feature mapping module, and the step (5) specifically comprises the following steps: inputting the second frame image into the backbone network FGet the complete characteristic diagram
Then, the coordinate information corresponding to the category in the second frame image is predicted
And the coordinate information P of the second frame image is retained in the feature map
Cutting the part corresponding to the position, inputting the two parts of features obtained by cutting into a feature mapping module respectively, and outputting 1 x 128-dimensional predicted shape matching features
And profile matching features
. Compared with the original output characteristic diagram, the characteristic diagram after mapping is integrated and compressed with effective characteristics, and is more efficient in measuring the similarity of the shape information of different positions.
The matching feature acquisition network realizes the functions of matching and re-identification, adopts triple loss during training, takes the same target image in a video sequence to form a positive sample pair, different target images to form a negative sample pair, and sets an anchor sample to obtain the shape matching feature
The positive sample has the shape matching characteristic of
The negative sample has the shape matching characteristic of
Then define
Comprises the following steps:
wherein i represents the ith training sample, and + represents that the original value is taken when the value is greater than 0 and 0 is taken when the value is less than 0, and alpha is the minimum interval between different samples.
(6) Predicting the coordinate information of the second frame image according to the coordinate information P of the second frame image
Predicted shape matching features
And profile matching features
Calculating a distance metric
The distance measure
: the calculation process of (2) is as follows:
wherein the content of the first and second substances,
coordinate information P representing the second frame image and coordinate information of the predicted second frame image
The ratio of the intersection of (a) to the union thereof.
Using Hungarian algorithm, coordinate information P of the second frame image is compared with coordinate information of the predicted second frame image
Matching corresponding targets; if the matching fails, continuously predicting a certain number of frames for the motion prediction result corresponding to the non-detection result until the matching is achieved in a certain frame, or continuously predicting the certain number of frames and then failing to achieve the matching with the detection result, suspending the track operation and not predicting any more; re-identifying the detection result without the matching of the motion prediction result, and measuring the Euclidean distance between the shape matching feature of the pause track T and the shape matching feature of the current detection result P
Setting a threshold value delta, restarting the track motion prediction if the distance is smaller than the threshold value, simultaneously associating the current matching detection results, and establishing a motion prediction model for a new target if all tracks are not smaller than the threshold value.
Coordinate information P for the second frame image and coordinate information of the predicted second frame image
In the matched result, recording the detection result in the matched result as an observed value D, and calculating Kalman gain by combining with a measurement noise covariance matrix R
And updating the motion state of the second frame image
Error matrix
(7) Inputting the characteristics of the second frame image I and the matched coordinate information P and predicted coordinate information P' into the tracking result correction network, and outputting correctionCoordinate information of multi-target tracking; outputting the amount of deviation of the tracking result correction frame with respect to the predicted position of the motion
And respectively representing the offset of the correction result relative to the xy coordinate of the center store of the motion prediction frame, the offset of the width and the offset of the height, thereby obtaining a more accurate tracking result.
The tracking correction network consists of 4 layers of 3 × 3 convolutional layers and fully-connected layers, the output is the offset prediction of the regression matching pair quantity × 4, and Smooth-L1Loss is adopted during training. Setting the true value of the tracking result as
The result of the motion prediction is
The network output offset is
Then, then
Is defined as:
smooth L1Loss is such that the difference between the true and predicted values is not too large when the difference is large, but small enough when the distance is small, so that the network learns a more stable offset regression capability. The final tracking result is
。
(8) And (5) sequentially repeating the steps (2) to (7) on the subsequent frame image until all the video frame images are tracked.
The tracking method further comprises the following steps: after the tracking task of the frame image is completed, the shape matching feature F of the tracking corresponding detection result is usedPUpdate each stripShape matching feature F of tracking trajectoryT：
；
Wherein the content of the first and second substances,
it is indicated that the learning rate is,
representing the shape-matching features of the trace before processing the current frame,
and showing the track shape matching characteristics after the track corresponding to the target shape information of the current frame is updated.
For the evaluation of the tracking result, it is calculated by the following formula:
wherein
Respectively the number of missed detection, false detection and false matching in t frames.
The multi-target tracking system based on the fusion detection technology is used for multi-target tracking of pedestrians, American group take-out electric vehicles, hungry take-out electric vehicles and non-take-out electric vehicles, and the number of the pedestrians, the American group take-out electric vehicles, the hungry take-out electric vehicles and the non-take-out electric vehicles in the image at the 1 st second is respectively 2, 1, 2 and 6; the number of missed detections is 0, 0, 0, 0; the false detection numbers are respectively 0, 0, 0 and 1; the number of error matches is 0, 0, 0, 1; the tracking result score was calculated to be 81.82%. In the image of 1.5 seconds, the numbers of pedestrians, beauty-group takeaway electric vehicles, hungry take-away electric vehicles and non-takeaway electric vehicles are respectively 3, 3, 2 and 6; the number of missed detections is 0, 0, 0 and 1 respectively; the false detection numbers are respectively 0, 0, 0 and 1; the number of error matches is 0, 0, 0, 1; the tracking result score was calculated to be 78.50%. The tracking scores of the images in other times are not described in detail. From the tracking result fraction, the deep convolutional neural network and the target detection network used in the invention provide the reliability of the result, and the Kalman filter model, the matching feature acquisition network and the tracking result correction network used in the invention can ensure that the matching and tracking results are more accurate and have excellent performance.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.
Claims (6)
1. A multi-target tracking system based on fusion detection technology, the system comprising: the system comprises a deep convolutional neural network, a target detection network, a Kalman filter model, a matching feature acquisition network and a tracking result correction network; the output end of the deep convolutional neural network is connected with the input end of a target detection network, the output end of the target detection network is respectively connected with the input end of a matched feature acquisition network and the input end of a Kalman filter model, the output end of the Kalman filter model is connected with the input end of the matched feature acquisition network, and the output end of the matched feature acquisition network is connected with the input end of a tracking result correction network.
2. The multi-target tracking system tracking method based on the fusion detection technology as claimed in claim 1, which is characterized by comprising the following steps:
(1) collecting video frame images;
(2) outputting a first frame image and a second frame image of the video frame image to a deep convolutional neural network to obtain the characteristics of each target in the first frame image and the second frame image;
(3) respectively inputting the characteristics of each target in the first frame image and the second frame image into a target detector network, outputting the confidence score, the category type, the category score and the coordinate information of each target in the first frame image and the second frame image, calculating the product of the confidence score and the category score of each target, and reserving the category type of the target with the product higher than a threshold value and the corresponding coordinate information;
(4) inputting the category type of the reserved first frame image and the corresponding coordinate information into a Kalman filter model for target tracking prediction, and predicting the coordinate information corresponding to the category type in the second frame image;
(5) inputting the second frame image, the coordinate information corresponding to the category type in the predicted second frame image and the reserved coordinate information of the second frame image into a matching feature acquisition network to obtain a predicted shape matching feature and a shape matching feature;
(6) calculating distance measurement according to the coordinate information of the second frame image, the coordinate information of the predicted second frame image, the predicted appearance matching characteristic and the appearance matching characteristic, and matching the coordinate information of the second frame image with a target corresponding to the coordinate information of the predicted second frame image by using a Hungarian algorithm;
(7) inputting the predicted shape matching feature and the shape matching feature of the second frame image obtained in the step (5) and the matched coordinate information and predicted coordinate information into a tracking result correction network, and outputting corrected coordinate information of multi-target tracking;
(8) and (5) sequentially repeating the steps (2) to (7) on the subsequent frame image until all the video frame images are tracked.
3. The tracking method of the multi-target tracking system based on the fusion detection technology as claimed in claim 2, wherein the matching feature acquisition network is composed of a backbone network and a feature mapping module, and the step (5) is specifically as follows: inputting the second frame image into a backbone network to obtain a complete feature map, cutting the position part of the coordinate information corresponding to the category type in the predicted second frame image and the reserved coordinate information of the second frame image in the feature map, respectively inputting the two cut parts of features into a feature mapping module, and outputting a 1 x 128-dimensional predicted shape matching feature and a shape matching feature.
4. The method as claimed in claim 2, wherein the distance metric is measured by a distance measuring device
: the calculation process of (2) is as follows:
wherein the content of the first and second substances,
coordinate information P representing the second frame image and coordinate information of the predicted second frame image
The ratio of the intersection of (a) to the union thereof,
in order to predict the shape-matching features,
is a shape matching feature.
5. The tracking method of the multi-target tracking system based on the fusion detection technology as claimed in claim 2, wherein in the step (6), if the matching fails, the motion prediction result is continuously predicted for a certain number of frames until the matching is achieved for a certain frame, or the motion prediction result is not matched with the detection result after the certain number of frames are continuously predicted, the track operation is suspended, and the prediction is not performed; if there is no detection result matching the motion prediction result, re-recognition is performed to determine whether to restart the old target or to newly initialize a new target.
6. The method of claim 2The tracking method of the multi-target tracking system based on the fusion detection technology is characterized by further comprising the following steps: after the tracking task of the frame image is completed, the shape matching feature F of the tracking corresponding detection result is usedPUpdating the shape matching feature F of each tracking trackT：
；
Wherein the content of the first and second substances,
it is indicated that the learning rate is,
representing the shape-matching features of the trace before processing the current frame,
and showing the track shape matching characteristics after the track corresponding to the target shape information of the current frame is updated.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110519994.8A CN112949615B (en) | 2021-05-13 | 2021-05-13 | Multi-target tracking system and method based on fusion detection technology |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110519994.8A CN112949615B (en) | 2021-05-13 | 2021-05-13 | Multi-target tracking system and method based on fusion detection technology |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112949615A true CN112949615A (en) | 2021-06-11 |
| CN112949615B CN112949615B (en) | 2021-08-17 |
Family
ID=76233798
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110519994.8A Active CN112949615B (en) | 2021-05-13 | 2021-05-13 | Multi-target tracking system and method based on fusion detection technology |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112949615B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115719368A (en) * | 2022-11-29 | 2023-02-28 | 上海船舶运输科学研究所有限公司 | Multi-target ship tracking method and system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107563313A (en) * | 2017-08-18 | 2018-01-09 | 北京航空航天大学 | Multiple target pedestrian detection and tracking based on deep learning |
| US20190073783A1 (en) * | 2006-01-04 | 2019-03-07 | Mobileye Vision Technologies Ltd. | Estimating distance to an object using a sequence of images recorded by a monocular camera |
| CN111127513A (en) * | 2019-12-02 | 2020-05-08 | 北京交通大学 | Multi-target tracking method |
| CN111739053A (en) * | 2019-03-21 | 2020-10-02 | 四川大学 | Online multi-pedestrian detection tracking method under complex scene |
| CN112288770A (en) * | 2020-09-25 | 2021-01-29 | 航天科工深圳（集团）有限公司 | Video real-time multi-target detection and tracking method and device based on deep learning |
-
2021
- 2021-05-13 CN CN202110519994.8A patent/CN112949615B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190073783A1 (en) * | 2006-01-04 | 2019-03-07 | Mobileye Vision Technologies Ltd. | Estimating distance to an object using a sequence of images recorded by a monocular camera |
| CN107563313A (en) * | 2017-08-18 | 2018-01-09 | 北京航空航天大学 | Multiple target pedestrian detection and tracking based on deep learning |
| CN111739053A (en) * | 2019-03-21 | 2020-10-02 | 四川大学 | Online multi-pedestrian detection tracking method under complex scene |
| CN111127513A (en) * | 2019-12-02 | 2020-05-08 | 北京交通大学 | Multi-target tracking method |
| CN112288770A (en) * | 2020-09-25 | 2021-01-29 | 航天科工深圳（集团）有限公司 | Video real-time multi-target detection and tracking method and device based on deep learning |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115719368A (en) * | 2022-11-29 | 2023-02-28 | 上海船舶运输科学研究所有限公司 | Multi-target ship tracking method and system |
| CN115719368B (en) * | 2022-11-29 | 2024-05-17 | 上海船舶运输科学研究所有限公司 | Multi-target ship tracking method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112949615B (en) | 2021-08-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Dedeoğlu et al. | Silhouette-based method for object classification and human action recognition in video | |
| Mallikarjuna et al. | Traffic data collection under mixed traffic conditions using video image processing | |
| Meuter et al. | A decision fusion and reasoning module for a traffic sign recognition system | |
| CN111310583A (en) | Vehicle abnormal behavior identification method based on improved long-term and short-term memory network | |
| Lee et al. | Learning discriminative appearance models for online multi-object tracking with appearance discriminability measures | |
| CN106570490B (en) | A kind of pedestrian's method for real time tracking based on quick clustering | |
| CN106778712A (en) | A kind of multi-target detection and tracking method | |
| Dai et al. | Instance segmentation enabled hybrid data association and discriminative hashing for online multi-object tracking | |
| CN110781785A (en) | Traffic scene pedestrian detection method improved based on fast RCNN algorithm | |
| CN113256690B (en) | Pedestrian multi-target tracking method based on video monitoring | |
| CN112738470B (en) | Method for detecting parking in highway tunnel | |
| CN111882586A (en) | Multi-actor target tracking method oriented to theater environment | |
| Liu et al. | Dynamic RGB-D SLAM based on static probability and observation number | |
| Li et al. | Bi-directional dense traffic counting based on spatio-temporal counting feature and counting-LSTM network | |
| CN112232240A (en) | Road sprinkled object detection and identification method based on optimized intersection-to-parallel ratio function | |
| Chen et al. | A video-based method with strong-robustness for vehicle detection and classification based on static appearance features and motion features | |
| CN112949615B (en) | Multi-target tracking system and method based on fusion detection technology | |
| CN114820765A (en) | Image recognition method and device, electronic equipment and computer readable storage medium | |
| Abdullah et al. | Vehicle counting using deep learning models: a comparative study | |
| CN117011341A (en) | Vehicle track detection method and system based on target tracking | |
| Lafuente-Arroyo et al. | Road sign tracking with a predictive filter solution | |
| Dong et al. | An automatic object detection and tracking method based on video surveillance | |
| CN108830182B (en) | Lane line detection method based on cascade convolution neural network | |
| Buch et al. | Urban vehicle tracking using a combined 3D model detector and classifier | |
| CN112528937A (en) | Method for detecting starting and stopping of video pumping unit |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |