CN111223121A - Multi-target track prediction method based on collision-free LSTM - Google Patents
Multi-target track prediction method based on collision-free LSTM Download PDFInfo
- Publication number
- CN111223121A CN111223121A CN202010031249.4A CN202010031249A CN111223121A CN 111223121 A CN111223121 A CN 111223121A CN 202010031249 A CN202010031249 A CN 202010031249A CN 111223121 A CN111223121 A CN 111223121A
- Authority
- CN
- China
- Prior art keywords
- target
- lstm
- collision
- free
- repulsive force
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 30
- 230000003321 amplification Effects 0.000 claims description 11
- 238000001514 detection method Methods 0.000 claims description 11
- 238000003199 nucleic acid amplification method Methods 0.000 claims description 11
- 238000012544 monitoring process Methods 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 6
- 238000012360 testing method Methods 0.000 claims description 4
- 238000012549 training Methods 0.000 claims description 4
- 230000009471 action Effects 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 2
- 238000012545 processing Methods 0.000 claims description 2
- 238000011176 pooling Methods 0.000 abstract description 4
- 238000013135 deep learning Methods 0.000 abstract 1
- 239000006185 dispersion Substances 0.000 abstract 1
- 238000010586 diagram Methods 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 230000006870 function Effects 0.000 description 4
- 230000003993 interaction Effects 0.000 description 3
- 206010000117 Abnormal behaviour Diseases 0.000 description 2
- 230000000875 corresponding effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000009792 diffusion process Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30232—Surveillance
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30241—Trajectory
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a multi-target track prediction method based on Collision-Free LSTM (CF-LSTM). A multi-target track prediction model based on deep learning is provided, a repulsive force model based on a target distance relation is established by utilizing track sequence data, a repulsive force sharing pool (dispersion Pooling) is further designed and introduced into an LSTM, and a collision-free track prediction network based on data driving is successfully constructed. The model can adapt to different scenes, learn factors influencing the track direction of the target and predict the tracks of a plurality of targets in a group.
Description
Technical Field
The invention belongs to the technical field of target track prediction, and particularly relates to a collision-free LSTM-based multi-target track prediction method.
Background
As a human, the human has an intuitive navigation capability, and can be considered as an intelligent object dominated by the human whether pedestrians, bicycles or vehicles. Studies have shown that human mobility is highly predictable and that their behaviour patterns can be predicted given sufficient observation information. When people walk in a crowded public space, such as a train terminal, a shopping center or a city center, they are subject to a large amount of (unprofitable) common sense and regulations, comply with social practices, traffic regulations, and the like. For example, people may keep a distance from surrounding objects for safety and personal space considerations, considering where to move next. Therefore, the influenced factors of the trajectory of the human-dominated agent target are very complex, and all the factors are difficult to be abstractly modeled.
The multi-target track prediction detects the abnormal behavior of the group targets by predicting the group multi-tracks, can timely and accurately send out alarm signals to emergencies, and is convenient for rapidly developing corresponding actions. In areas with heavy activities or dense population, the detection of abnormal behaviors in the population is also an important means for ensuring safety. How to utilize stable monitoring images to carry out predictive analysis on multiple tracks of group targets is an important problem currently faced.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention aims to provide a multi-target track prediction method based on Collision-Free LSTMs, wherein a Collision-Free (CF-LSTM) sequence learning model is utilized, each LSTM in the model represents a target track, the whole LSTMs network predicts the track coordinates of the targets by learning the relation hidden state among the targets, and different distances among the targets have different influence degrees (namely, the target closest to the target has larger influence than other targets). Meanwhile, the invention designs a repulsive force pool (R-Pooling), and each LSTM node receives information from the repulsive force sharing pool, and the information represents the repulsive force influence of adjacent targets in a certain spatial range. The method can adapt to different scenes, learn factors influencing the track direction of the target and predict the tracks of a plurality of targets in a group.
In order to achieve the purpose, the invention adopts the technical scheme that:
a multi-target track prediction method based on collision-free LSTM comprises the following steps:
step 1: obtaining a motion track of a target from a monitoring video;
step 2: calculating a hidden state tensor influencing each target direction through a repulsive force sharing pool;
and step 3: and (4) performing position prediction on multiple targets by using a collision-free LSTM model.
In the step 1, the motion trail of the target in the monitoring picture is obtained through the monitoring video, and the specific method comprises the following steps:
through the preprocessing of the monitoring video, a certain track in the obtained monitoring video is a coordinate sequence arranged according to time, and the number of continuous points of the track in monitoring is the length of the sequence. The video is subjected to frame extraction processing, for example, every 15 frames, a moving object detection result concerned by a current frame is intercepted, then a moving object in the frame is detected by using an object detection algorithm, a geometric center of the detected moving object is used as an object center point, and a moving coordinate of the moving object center point is recorded, so that a moving track of the object is obtained.
In the step 2, repulsive force models are used for calculating repulsive force among targets of different types, and then a repulsive force sharing pool is used for calculating hidden state tensors influencing each target direction.
The repulsive force between the different classes of targets is calculated as follows:
determining the space coordinates of the targets, setting the diameters of the targets of different types, setting the amplification factors and distance adjustment factors of the targets of different types, and firstly calculating the Euclidean distance between the targets
Further, the repulsive force model is used to calculate the repulsive force Ri,j(Si,j)＝ai,jexp(Si,j/Di,j)ai,j＞0,Di,j> 0, wherein (x)i,yi) And (x)j,yj) Is the coordinates of the target i targets j, parameter ai,jFor amplification of the magnitude of the repulsive force, parameter Di,jIs the adjustment factor of the distance of action of the repulsive force.
It should be noted that the amplification factor and the adjustment factor are different between different classes of targets, such as: the human-to-human amplification factor and the adjustment factor and the human-to-bicycle amplification factor and the adjustment factor are different in magnitude. The invention sets three types of targets, human, bicycle and automobile respectively.
The calculation method of the hidden state tensor influencing each target direction is as follows:
determining a predicted target, setting the radius of a repulsive force sharing pool with the length of r and taking the predicted target as the center, dividing N fan-shaped areas, wherein the number of the fan-shaped areas is the direction dimension, calculating the repulsive force of each target in the range of the predicted target and the repulsive force sharing pool by utilizing a repulsive force model, and calculating each fan-shaped area, namely the shared hidden state tensor of the predicted target influenced by different methods.
In step 3, the method for predicting the positions of multiple targets comprises the following steps:
and (3) predicting a plurality of spatial positions of the motion trail of the target by the spatial positions of the motion trail of the target obtained in the step (1), namely, the input sequence of the collision-free LSTM model is the spatial positions of all targets in a certain space, and the output sequence is the position of a track point in the future of the predicted target.
The prediction calculation method for the multiple spatial positions of the target track comprises the following steps:
the track distribution conforms to binary Gaussian distribution, and the track position at the moment t can be predicted through the hidden state of the first t-1 times.
The parameters of the collision-free LSTM model are optimized as follows:
given the real trajectory spatial coordinate values and the parameter set of the target, the collision-free LSTM network optimizes the model parameters by minimizing the negative log-likelihood estimates.
Inputting the spatial position information of the plurality of target tracks into a collision-free LSTM model, and further outputting the spatial position information of the plurality of target tracks:
during training, setting direction dimensions, shared pool space radius, dimensions of a hidden state of a collision-free LSTM model and the number of observed track values by using the description of the existing data set, predicting tracks of future frame numbers, optimizing a target function by using a gradient descent method, updating weights by using a back propagation algorithm, and further training a collision-free LSTM network;
during testing, a plurality of target track space coordinate positions obtained from the observation frame number in a test video are input into the collision-free LSTM, and the space position output by the network is the predicted track information.
Compared with the prior art, the invention has the beneficial effects that: the model can adapt to different scenes, learn factors influencing the target track direction, acquire the hidden state of interaction between targets and predict the tracks of a plurality of targets.
Drawings
FIG. 1 is a schematic diagram of a collision-free LSTM model.
Fig. 2 is a schematic diagram of a process of extracting a moving target track, wherein: (a) an initial target detection map, (b) a 15 frame later target detection map, (c) a 30 frame later target detection map, (d) a 45 frame target detection map, (e) a 60 frame later target detection map, (f) a 75 frame later target detection map, and (g) a moving target track map.
FIG. 3 is a schematic diagram of the repulsive force sharing pool model operation.
Fig. 4 is a trajectory prediction network decoding process.
Fig. 5 is a schematic diagram of a repulsive force model.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.
The invention discloses a Collision-Free LSTM-based multi-target track prediction method, which comprises the steps of forming track sequence data in a period of time by utilizing continuous recording of the space coordinate position of each target, establishing a repulsive force model based on a target distance relation, designing a repulsive force sharing pool (replay), trying to be introduced into an LSTM network, and successfully establishing a sequence learning-based track prediction network (Collision-Free LSTM, CF-LSTM). The model can adapt to different scenes, learn factors influencing the target track direction, acquire the hidden state of interaction between targets and predict the tracks of a plurality of targets.
As shown in FIG. 1, the left diagram shows a detailed description of the repulsive force sharing pool (R-Pooling) with respect to the target 2, the target 2 being in the range of radius RDifferent positions have different repulsive effects on the target 2. The right side shows that LSTMs are connected to each target in a certain range by a repulsive force sharing pool (R-Pooling), and the hidden states of target 1 and target 3 at a certain time are h1And h3Calculating the repulsive force influence value H on the target 2 through a repulsive force sharing pool2。
The specific method of the invention is as follows:
1. target motion trajectory extraction
As shown in fig. 2 (a) - (f), for every 15 frames of the surveillance video, capturing the detection result of the moving object concerned by the current frame, detecting the image object by using a Mask R-CNN model, taking the geometric center of the detected moving object as the object center point, and recording the moving coordinates of the center point of the moving object to obtain the moving track of the object (fig. 2 (g)). The trajectory may be expressed as equation (1).
Wherein the content of the first and second substances,
indicating the coordinates of the ith target at time t. The observed trajectory is from time 1 to TobsIs the task of predicting the position of the target from time Tobs+1 to Tpred. This can be considered a sequence generation problem where the input sequence is the spatial positions of all targets in a space and the output sequence is the position of the track points in the future of the predicted target.
2. Repulsive force sharing pool calculation hidden state tensor
(1) Repulsive force calculation model
Different objects on the way they travel will intentionally avoid collisions with nearby objects in the surroundings. Therefore, the social force model can be further simplified, and a simple and effective target motion direction model is established according to repulsive force generated by the distance between the concerned target and the surrounding target without considering the speed factor of the target. As shown in fig. 5, rowsRepulsion force model (x)i,yi)、(xj,yj) And (x)k,yk) Respectively representing the space position coordinates of object i, object j and object k, the distance S between the objectsi,jSee equation (2).
The repulsive force calculation process is shown in the formula (3),
Ri,j(Si,j)＝ai,jexp(Si,j/Di,j)ai,j＞0,Di,j＞0 (3)
wherein the parameter ai,jFor amplification of the magnitude of the repulsive force, parameter Di,jIs the adjustment factor of the distance of action of the repulsive force.
It should be noted that the amplification factor and the adjustment factor are different between different classes of targets, such as: the human-to-human amplification factor and the adjustment factor and the human-to-bicycle amplification factor and the adjustment factor are different in magnitude. This patent sets for three types of targets, human, bicycle and automobile respectively.
(2) Repulsive force sharing pool
The proposed repulsive force sharing pool model (diffusion), can gather interaction information between targets in different directions within a certain range. Through a repulsive force model, the shared pool can calculate a repulsive force hidden state tensor, and the tensor information expresses the hidden state of mutual influence between adjacent targets in different directions. And finally, outputting hidden state information of the node influencing the target motion trail through LSTM node operation. As shown in fig. 3 in particular, a certain spatial area around the target 1 in the repulsive force sharing pool is divided into N sector-shaped regions in total. In the regions (1,2) and the regions (1,3), objects 3, 4, and 2 are collected, respectively. In step 2, targets of different distances have different repulsive forces on target 1. The last step is to display the repulsive force hidden tensor H of the object 1 in a different direction1。
First, the direction dimension is set to be N, and the hidden state dimension of LSTM is set to be Q.
This indicates that the repulsive force of the target i at time t corresponding to the azimuth n is hidden, as shown in equation (4). Then, a repulsive implicit state tensor is constructed in the whole direction dimension N, as shown in the formula (5),
wherein R (S)i，j) The repulsive force between the target i and the target j is expressed, and the formula is calculated and shown in the formula (5);
indicating the LSTM hidden state of target j at time t-1; (i, n) represents a region of the target i direction n; function 1i,n[x,y]Judging whether the space position (x, y) is in the area (i, n); mi,nRepresenting a set of objects within region (i, n).
Mapping spatial coordinates (x, y) to r using a function phi (·)i tMapping the repulsive force tensor to
The encoding process of the LSTM network at time t is as follows:
where φ (-) is a nonlinear transformation function ReLU; wr and Wc are conversion weights; wlRepresents the LSTM sharing parameter;
3. location prediction process
By the hidden state of the first t-1 time, the prediction can be realized
The track position of (2). Assuming that the trajectory distribution conforms to a binary gaussian distribution, the mean is expressed as:
standard deviation of
Standard deviation of
A weight matrix W of 5 xQ predicting these parametersp. The method comprises the following specific steps:
the CF-LSTM optimizes the model parameters by minimizing the negative log-likelihood estimates. Equation (11) represents the trajectory of the prediction target i. The specific decoding is shown in fig. 4, which shows a track prediction, as shown in fig. 5.
Claims (9)
1. A multi-target track prediction method based on collision-free LSTM is characterized by comprising the following steps:
step 1: obtaining a motion track of a target from a monitoring video;
step 2: calculating a hidden state tensor influencing each target direction through a repulsive force sharing pool;
and step 3: and (4) performing position prediction on multiple targets by using a collision-free LSTM model.
2. The method for predicting multiple target tracks based on collision-free LSTM as claimed in claim 1, wherein in step 1, the motion track of the target in the monitored picture is obtained by the monitoring video, and the specific method is as follows:
and performing frame extraction processing on the video, detecting a moving target in the frame by using a target detection algorithm, taking the geometric center of the detected moving target as a target central point, and recording the moving coordinate of the moving target central point to obtain the motion track of the target.
3. The collision-free LSTM-based multi-target trajectory prediction method of claim 1, wherein in step 2, repulsive forces between targets of different classes are calculated using a repulsive force model, and then hidden state tensors affecting each target direction are calculated using a repulsive force sharing pool.
4. The collision-free LSTM-based multi-target trajectory prediction method of claim 3, wherein the repulsive force between the targets of different classes is calculated as follows:
determining the space coordinates of the targets, setting the diameters of the targets of different types, setting the amplification factors and distance adjustment factors of the targets of different types, and firstly calculating the Euclidean distance between the targets
Further, the repulsive force model is used to calculate the repulsive force Ri,j(Si,j)＝ai,jexp(Si,j/Di,j)ai,j＞0,Di,j> 0, wherein (x)i,yi) And (x)j,yj) Is the coordinates of the target i targets j, parameter ai,jFor amplification of the magnitude of the repulsive force, parameter Di,jIs the adjustment factor of the distance of action of the repulsive force.
5. The collision-free LSTM-based multi-target trajectory prediction method of claim 3, wherein the hidden state tensor affecting each target direction is computed as follows:
determining a predicted target, setting the radius of a repulsive force sharing pool with the length of r and taking the predicted target as the center, dividing N fan-shaped areas, wherein the number of the fan-shaped areas is the direction dimension, calculating the repulsive force of each target in the range of the predicted target and the repulsive force sharing pool by utilizing a repulsive force model, and calculating each fan-shaped area, namely the shared hidden state tensor of the predicted target influenced by different methods.
6. The collision-free LSTM-based multi-target track prediction method of claim 1, wherein in step 3, the position prediction method for the multiple targets is as follows:
and (3) predicting a plurality of spatial positions of the motion trail of the target by the spatial positions of the motion trail of the target obtained in the step (1), namely, the input sequence of the collision-free LSTM model is the spatial positions of all targets in a certain space, and the output sequence is the position of a track point in the future of the predicted target.
7. The collision-free LSTM-based multi-target trajectory prediction method of claim 6, wherein the prediction calculation method for the multiple spatial positions of the target trajectory is as follows:
the track distribution conforms to binary Gaussian distribution, and the track position at the moment t can be predicted through the hidden state of the first t-1 times.
8. The collision-free LSTM-based multi-target trajectory prediction method of claim 6, wherein the parameters of the collision-free LSTM model are optimized as follows:
given the real trajectory spatial coordinate values and the parameter set of the target, the collision-free LSTM network optimizes the model parameters by minimizing the negative log-likelihood estimates.
9. The collision-free LSTM-based multi-target track prediction method of claim 6, wherein the spatial position information of the plurality of target tracks is input into the collision-free LSTM model, and further the spatial position information of the plurality of target tracks is output:
during training, setting direction dimensions, shared pool space radius, dimensions of a hidden state of a collision-free LSTM model and the number of observed track values by using the description of the existing data set, predicting tracks of future frame numbers, optimizing a target function by using a gradient descent method, updating weights by using a back propagation algorithm, and further training a collision-free LSTM network;
during testing, a plurality of target track space coordinate positions obtained from the observation frame number in a test video are input into the collision-free LSTM, and the space position output by the network is the predicted track information.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010031249.4A CN111223121A (en) | 2020-01-13 | 2020-01-13 | Multi-target track prediction method based on collision-free LSTM |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010031249.4A CN111223121A (en) | 2020-01-13 | 2020-01-13 | Multi-target track prediction method based on collision-free LSTM |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111223121A true CN111223121A (en) | 2020-06-02 |
Family
ID=70831129
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010031249.4A Pending CN111223121A (en) | 2020-01-13 | 2020-01-13 | Multi-target track prediction method based on collision-free LSTM |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111223121A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112597922A (en) * | 2020-12-28 | 2021-04-02 | 深圳市普汇智联科技有限公司 | Multi-target track prediction method based on collision-free LSTM |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107492113A (en) * | 2017-06-01 | 2017-12-19 | 南京行者易智能交通科技有限公司 | A kind of moving object in video sequences position prediction model training method, position predicting method and trajectory predictions method |
| CN108564118A (en) * | 2018-03-30 | 2018-09-21 | 陕西师范大学 | Crowd scene pedestrian track prediction technique based on social affinity shot and long term memory network model |
| CN110135314A (en) * | 2019-05-07 | 2019-08-16 | 电子科技大学 | A kind of multi-object tracking method based on depth Trajectory prediction |
-
2020
- 2020-01-13 CN CN202010031249.4A patent/CN111223121A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107492113A (en) * | 2017-06-01 | 2017-12-19 | 南京行者易智能交通科技有限公司 | A kind of moving object in video sequences position prediction model training method, position predicting method and trajectory predictions method |
| CN108564118A (en) * | 2018-03-30 | 2018-09-21 | 陕西师范大学 | Crowd scene pedestrian track prediction technique based on social affinity shot and long term memory network model |
| CN110135314A (en) * | 2019-05-07 | 2019-08-16 | 电子科技大学 | A kind of multi-object tracking method based on depth Trajectory prediction |
Non-Patent Citations (1)
| Title |
|---|
| KAIPING XU等: "Collision-Free LSTM for Human Trajectory Prediction", 《INTERNATIONAL CONFERENCE ON MULTIMEDIA MODELING》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112597922A (en) * | 2020-12-28 | 2021-04-02 | 深圳市普汇智联科技有限公司 | Multi-target track prediction method based on collision-free LSTM |
| CN112597922B (en) * | 2020-12-28 | 2024-03-26 | 深圳市普汇智联科技有限公司 | Multi-target track prediction method based on collision-free LSTM |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yao et al. | Unsupervised traffic accident detection in first-person videos | |
| CN109961019B (en) | Space-time behavior detection method | |
| KR102129893B1 (en) | Ship tracking method and system based on deep learning network and average movement | |
| Xue et al. | SS-LSTM: A hierarchical LSTM model for pedestrian trajectory prediction | |
| CN110737968B (en) | Crowd trajectory prediction method and system based on deep convolutional long and short memory network | |
| US7613569B2 (en) | Crash prediction network with visual input for vehicle | |
| Nasernejad et al. | Modeling pedestrian behavior in pedestrian-vehicle near misses: A continuous Gaussian Process Inverse Reinforcement Learning (GP-IRL) approach | |
| Wang et al. | [Retracted] Innovative Research of Trajectory Prediction Algorithm Based on Deep Learning in Car Network Collision Detection and Early Warning System | |
| CN104318578A (en) | Video image analyzing method and system | |
| CN113642403B (en) | Crowd abnormal intelligent safety detection system based on edge calculation | |
| CN114267082B (en) | Bridge side falling behavior identification method based on depth understanding | |
| CN112767451A (en) | Crowd distribution prediction method and system based on double-current convolutional neural network | |
| Khosravi et al. | Crowd emotion prediction for human-vehicle interaction through modified transfer learning and fuzzy logic ranking | |
| CN113436217A (en) | Unmanned vehicle environment detection method based on deep learning | |
| CN116090333A (en) | Urban public space disaster modeling and preventing system based on perception blind area estimation | |
| CN115376103A (en) | Pedestrian trajectory prediction method based on space-time diagram attention network | |
| CN111223121A (en) | Multi-target track prediction method based on collision-free LSTM | |
| CN105469054A (en) | Model construction method of normal behaviors and detection method of abnormal behaviors | |
| CN107016358A (en) | Intermediate density scene microcommunity real-time detection method | |
| Liu et al. | A video drowning detection device based on underwater computer vision | |
| CN115979250B (en) | Positioning method based on UWB module, semantic map and visual information | |
| CN116501820A (en) | Vehicle track prediction method, device, equipment and storage medium | |
| CN116229341A (en) | Method and system for analyzing and alarming suspicious behaviors in video monitoring among electrons | |
| CN111881863B (en) | Regional group abnormal behavior detection method | |
| Hao et al. | Group abnormal behaviour detection algorithm based on global optical flow |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200602 |
|
| RJ01 | Rejection of invention patent application after publication |