CN116681722A - Traffic accident detection method based on isolated forest algorithm and target tracking - Google Patents

Traffic accident detection method based on isolated forest algorithm and target tracking Download PDF

Info

Publication number
CN116681722A
CN116681722A CN202310596197.9A CN202310596197A CN116681722A CN 116681722 A CN116681722 A CN 116681722A CN 202310596197 A CN202310596197 A CN 202310596197A CN 116681722 A CN116681722 A CN 116681722A
Authority
CN
China
Prior art keywords
vehicle
abnormal
vehicles
score
motion
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
Application number
CN202310596197.9A
Other languages
Chinese (zh)
Inventor
宋彬
王帅
秦浩
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Original Assignee
Xidian University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xidian University filed Critical Xidian University
Priority to CN202310596197.9A priority Critical patent/CN116681722A/en
Publication of CN116681722A publication Critical patent/CN116681722A/en
Pending legal-status Critical Current

Links

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
    • G06T7/207—Analysis of motion for motion estimation over a hierarchy of resolutions
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/70—Determining position or orientation of objects or cameras
    • G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
    • 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
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30196—Human being; Person
    • 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
    • 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
    • G06T2207/30248—Vehicle exterior or interior
    • G06T2207/30252—Vehicle exterior; Vicinity of vehicle
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/08—Detecting or categorising vehicles
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Traffic Control Systems (AREA)

Abstract

The invention discloses a traffic accident detection method based on an isolated forest algorithm and target tracking, which aims at the problem that the traffic accident detection effect still needs to be improved in the prior art. The invention adopts a target tracking algorithm to acquire the motion trail of each vehicle and pedestrian in the traffic video; calculating the motion index of each vehicle according to the motion trail of each vehicle; screening abnormal vehicles by using an isolated forest algorithm; providing a scoring mechanism for accumulating abnormal scores for each abnormal vehicle after screening; an attention redistribution mechanism is provided, and the attention coefficient obtained by the mechanism is used for weighting abnormal parts of all vehicles so as to highlight accident vehicles in all abnormal vehicles at the same moment; the video is divided into a plurality of discrimination periods, and in each discrimination period, the score efficiency of each abnormal vehicle is calculated from the accumulated abnormal score, and the accident vehicle is determined from the score efficiency. The invention can accurately and efficiently detect traffic accidents in traffic videos and accurately position accident vehicles.

Description

Traffic accident detection method based on isolated forest algorithm and target tracking
Technical Field
The invention relates to the technical field of image processing, in particular to a traffic accident detection method based on an isolated forest algorithm and target tracking.
Background
With the continuous acceleration of the urban process in the modern society, the urban traffic is more and more large in scale, traffic accidents are more frequent, and the demands of society on intelligent traffic accident recognition are increasingly urgent.
In the past decades, the event detection problem has raised the widespread interest of numerous researchers at home and abroad, and traffic accident detection algorithms based on methods such as pattern recognition, neural networks, support vector machines, fuzzy logic, kalman filtering, time sequence analysis and the like have achieved different degrees of results.
The california algorithm is an early highway traffic detection algorithm that discriminates between possible sudden traffic events by comparing traffic flow parameter data between adjacent detection points. Subsequently, standard deviation algorithms, double exponential smoothing algorithms, bayesian algorithms appear again. The McMaster algorithm based on mutation theory takes frequent congestion caused by excessive traffic demand as an object for analysis and judgment for the first time. In early traffic event detection algorithms, the California algorithm and the McMaster algorithm have obvious advantages in terms of accuracy, instantaneity and the like, are considered to be two classical algorithms with the best comprehensive performance, and are generally used as targets for evaluation of other algorithms. After the 90 s of the 20 th century, a series of artificial intelligence algorithms are gradually emerging in the automatic detection of traffic events. Among them, chew et al propose an artificial neural network structural model, which achieves a good detection effect in traffic event detection. Abdulhai applies probabilistic neural networks to emergency detection. Fang Yuan and Rucy Long Cheu propose a traffic event detection algorithm based on a support vector machine. Shawrya Agarwal proposes to detect traffic events using a hybrid model of wavelet transformation and logistic regression. The early AID algorithm becomes a classical algorithm for detecting highway traffic events based on pattern recognition and mutation theory. The later artificial intelligent AID algorithm is based on the innovation of the emerging theory and hardware technology, and is more advantageous in detection effect and detection performance. However, the current traffic accident detection algorithm does not achieve the ideal effect, so that the core algorithm needs to be studied in depth.
Disclosure of Invention
Aiming at the problem that the traffic accident detection effect still needs to be improved in the prior art, the invention provides the traffic accident detection method based on the isolated forest algorithm and the target tracking, which has more advantages in detection effect and performance.
The technical scheme of the invention is that the traffic accident detection method based on the isolated forest algorithm and the target tracking is provided, which comprises the following steps: comprises the following steps of: for input traffic video data, firstly, acquiring the positions of vehicles and pedestrians in each frame of image by using a target tracking technology, and carrying out data association on the positions of the vehicles and pedestrians in different frames to obtain the motion trail of each vehicle and pedestrian in the video; step 2: calculating the motion index of each vehicle according to the motion trail of each vehicle; step 3: screening abnormal vehicles according to different motion indexes by using an isolated forest algorithm; step 4: according to the screening result of the isolated forest algorithm, a scoring mechanism is provided for accumulating abnormal scores of each abnormal vehicle; step 5: the attention redistribution mechanism is provided, the obtained attention coefficient weights the abnormal parts of all vehicles, and accident vehicles in all abnormal vehicles at the same moment are highlighted; step 6: the video is divided into a plurality of discrimination periods, and in each discrimination period, the score efficiency of each abnormal vehicle is calculated from the accumulated abnormal score, and the accident vehicle is determined from the score efficiency.
Preferably, in the step 1, traffic accident detection is performed based on single-mode data of the video, and a target tracking algorithm is adopted to obtain motion trajectories of vehicles and pedestrians in the video data, wherein for each vehicle and pedestrian in each frame of image, the specific form of the motion trajectories is as follows: center point pixel coordinates, detection frame pixel height, detection frame pixel width, track ID, and frame number.
Preferably, in the step 2, motion indexes of the vehicles are calculated according to motion trajectories of the vehicles, motion trajectory information of the vehicles and pedestrians is compressed in a time dimension, for multi-frame trajectory information corresponding to each second in a video, different vehicles and pedestrians are distinguished according to a trackID, only trajectory information of each vehicle and pedestrian when the first occurs is reserved, a minimum time unit after trajectory compression is called as time, and then speed, acceleration and course angle of the vehicles are calculated according to positions of the vehicles at a plurality of times respectively; the method comprises the following specific steps: different vehicles are distinguished by track ID, and the same vehicle is arranged at two adjacent moments t 1 、t 2 (t 1 ＜t 2 ) The two pieces of track information in (a) are respectively: t is t 1 Moment center point pixel abscissa x 1 Pixel ordinate y 1 Detecting the frame pixel height h 1 Detection frame pixel width w 1 ；t 2 Moment center point pixel abscissa x 2 Pixel ordinate y 2 Detecting the frame pixel height h 2 And a detection frame pixel width w 2 ，
The scale normalization coefficient is calculated as follows:
the calculation speed is as follows:
wherein v is t2 x At t 2 Normalized component speed, v of time x-axis direction t2 y At t 2 Normalized component speed, v of moment y-axis direction t2 At t 2 Time speed;
the acceleration is calculated as follows:
the calculated heading angle is as follows:
where σ is a very small positive number.
Preferably, in the step 3, abnormal vehicles are screened according to different motion indexes by using an isolated forest algorithm trained in advance, and the speeds, accelerations and heading angles of all vehicles at the same time are respectively sent into three isolated forest models to obtain the vehicles with abnormal speeds, accelerations and heading angles.
Preferably, in the step 4, according to the screening result of the isolated forest algorithm, a scoring mechanism is provided for accumulating abnormal scores for each abnormal vehicle, where the scoring mechanism includes: calculating motion index anomaly score and vehicle environment anomaly score for vehicles with motion indexes abnormal, judging whether the vehicle environment is abnormal, calculating vehicle environment anomaly score if abnormal, and averaging the motion index anomaly score and the vehicle environment anomaly score to obtain anomaly score;
the motion index anomaly score calculation formula is as follows:
S motion ＝α·S speed +β·S acc +γ·S θ ，(α+β+γ＝1),
in the above, alpha, beta and gamma are respectively abnormal speed weight, abnormal acceleration weight and abnormal course angle weight; s is S speed 、S acc 、S θ Respectively, a speed anomaly score, an acceleration anomaly score and a course angle anomaly score, and S is the time of abnormal vehicle speed speed 1, otherwise 0; s when the acceleration of the vehicle is abnormal acc 1, otherwise 0; s when the heading angle of the vehicle is abnormal θ 1, otherwise 0;
the vehicle environment abnormality judging mode is as follows: when people exist nearby the vehicle and the normalized distance of the people and the vehicle is smaller than 1, scoring is carried out, and the calculation formula of the normalized distance of the people and the vehicle is as follows:
x in the above person For the centre point pixel abscissa, y of a person person The ordinate of the pixel of the center point of the person, h person Human detection frame pixel height, w person Human detection frame pixel width, x car For the centre point pixel abscissa, y of the vehicle car Is the ordinate of the pixel of the central point of the vehicle, h car Detecting frame pixel height, w for vehicle car The width of the detection frame pixels of the vehicle;
for each abnormal vehicle, if the vehicle environment is not abnormal, namely the minimum value in the normalized distance between all pedestrians and the vehicles is greater than 1, the vehicle environment abnormality is divided into 0, otherwise, the vehicle environment abnormality is calculated according to the following formula:
S distance ＝1-d min ，
in the above, d min Normalizing the minimum value in the distances between all pedestrians and the vehicles;
the calculation formula of the abnormal score of the vehicle is as follows:
in the above, S motion S is the abnormal division of movement index distance Is an abnormal component of the vehicle environment.
Preferably, the attention is paid in step 5The distribution mechanism transfers the attention coefficient part of partial abnormal vehicles at the same moment to other abnormal vehicles, weights the abnormal parts of the vehicles by using the obtained attention coefficient, and highlights accident vehicles in all abnormal vehicles at the same moment; wherein the attention redistribution mechanism is: the attention coefficient of each abnormal vehicle before the attention transfer is 1, and the total attention value of each moment is equal to the number of abnormal vehicles at the moment; for each abnormal vehicle, inquiring the early warning times T before the current moment, and mapping the early warning times T into weights W according to the following formula t ：
Statistics of maximum anomaly score S of all anomaly vehicles at current moment max For each abnormal vehicle at the current moment, mapping its abnormal score S to weight W according to the following s ：
For each abnormal vehicle, setting the total number of abnormal vehicles at the current moment as N, and calculating the attention coefficient of each abnormal vehicle after the attention is redistributed according to the following formula:
preferably, in the step 6, the video is divided into a plurality of discrimination periods, in each discrimination period, the score efficiency of each abnormal vehicle is calculated according to the accumulated abnormal score, and the accident vehicle is determined according to the score efficiency, wherein different abnormal vehicles are distinguished by track ID, the early warning times of each abnormal vehicle are accumulated according to whether the score efficiency of each abnormal vehicle exceeds the early warning threshold value, the abnormal vehicle with the score efficiency exceeding the alarm threshold value is determined as the accident vehicle, and the calculation formula of the score efficiency is as follows:
wherein L is the total time number of the continuous frames, W i j And S is i j The attention coefficient and the abnormality score of the ith abnormal vehicle at the jth moment are respectively.
Compared with the prior art, the traffic accident detection method based on the isolated forest algorithm and the target tracking has the following advantages: according to the method, each motion index of the vehicle is calculated based on the motion trail of the vehicle obtained by the target tracking algorithm, the vehicle in the video can be rapidly and accurately positioned and distinguished, only the single-mode data of the image is used as the algorithm input, the operation efficiency of the algorithm can be effectively improved, and the hardware cost is reduced. The isolated forest algorithm is used for screening and scoring vehicles with abnormal motion indexes, and compared with the method for directly judging whether the accident is abnormal according to the motion indexes or not, the method for judging the outlier is introduced to detect the abnormality, so that false detection caused by special conditions such as red lights, traffic jam and the like can be effectively avoided. In addition, a attention redistribution mechanism is introduced to highlight accident vehicles in abnormal vehicles, the mechanism can effectively increase the abnormal gap between the accident vehicles and other abnormal vehicles, reduce false detection caused by the abnormal surrounding vehicle movement indexes caused by the accident vehicles, and improve the overall performance of the algorithm. Finally, accident judgment is carried out by calculating the scoring efficiency of the vehicle in a section of continuous frames, and false detection caused by occasional factors such as temporary braking of a driver, errors of the algorithm pre-step and the like in a short time can be reduced by adopting a sectional judgment mode.
The invention can accurately and efficiently detect traffic accidents in traffic videos, accurately position the accident vehicles in images, and greatly improve the practicality of the invention by virtue of the practicality and fine granularity positioning capability, so that users can quickly and accurately know road conditions.
Drawings
FIG. 1 is a schematic workflow diagram of the present invention;
FIG. 2 is a schematic diagram of the working steps of the present invention;
FIG. 3 is a schematic diagram of the structure of the attention reassignment mechanism in the present invention;
FIG. 4 is a schematic illustration of a broken line of a vehicle anomaly without using a attention redistribution mechanism in the present invention;
FIG. 5 is a schematic illustration of a broken line of a vehicle anomaly using a attention redistribution mechanism in the present invention;
FIG. 6 is a schematic view of the detection result in the present invention.
Detailed Description
In order that those skilled in the art will better understand the present invention, a technical solution in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in which it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the present invention without making any inventive effort, shall fall within the scope of the present invention.
The traffic accident detection method based on the isolated forest algorithm and the target tracking is further described below with reference to the accompanying drawings and the specific embodiments: in this embodiment, as shown in fig. 1 and 2, step 1: for input traffic video data, firstly, acquiring the positions of vehicles and pedestrians in each frame of image by using a target tracking technology, and carrying out data association on the positions of the vehicles and pedestrians in different frames to obtain the motion trail of each vehicle and pedestrian in the video. The specific form of the vehicle motion trail is as follows: the pixel coordinates of the center point, the pixel height of the detection frame, the pixel width of the detection frame, the track ID of the vehicle and the frame number of each vehicle in each frame of image; and acquiring the positions of the vehicle and the pedestrian of each frame in the video by adopting a YOLOX target detection algorithm. The YOLOX target detection algorithm adopts an anchor-free mechanism in a sample matching algorithm in a training process, so that the detection performance of a small target object is improved pertinently, and the small target object in video data is a target far away from image acquisition equipment in an urban traffic scene, so that the working distance of the algorithm can be effectively increased by adopting the target detection algorithm which is well performed on the small target object, and further the hardware cost is reduced. It should be noted that the selection of the target detection algorithm is not limited in the present invention.
In the embodiment, a ByteTrack multi-target tracking algorithm is adopted to carry out data association on the vehicle positions in different frames, so that the motion trail of each vehicle and each pedestrian in the video is obtained. Under the urban traffic scene, the invention has more small targets and serious shielding among targets, so that the target detection algorithm tends to output a large number of low-resolution detection frames, and the ReID characteristics among adjacent targets are similar. The ByteTrack multi-target tracking algorithm gives up the ReID feature of the target and only uses the motion model for data association. And compared with other target tracking algorithms, the low-score detection frame is not directly abandoned, and the low-score detection frame is utilized for secondary track matching. It should be noted that the present invention does not limit the selection of the target tracking algorithm.
Alternatively, bdd k data sets may be employed to train the target detection algorithm model, the target tracking algorithm model. bdd100k data set is a large-scale, diverse automated driving data set published by the AI laboratory at the university of bernoulli, and contains 10 ten thousand images, a total of about 184 ten thousand marker boxes for 10 categories.
In this embodiment, a YOLOX target detection algorithm is adopted, a YOLOX-m pre-training model provided by authorities is used for initializing network model parameters, a bdd100k data set and a self-built data set training model are respectively used, and training periods are 300 rounds. And embedding the trained yolox-m model into a ByteTrack multi-target tracking algorithm to acquire the motion trail of each vehicle in the video.
Step 2: and calculating the motion index according to the motion trail of each vehicle. In order to ensure accurate calculation of motion indexes and reduce algorithm complexity, firstly, motion track information of vehicles and pedestrians is compressed in a time dimension, a minimum time unit after track compression is called time, and then, according to positions of each vehicle at a plurality of times, speed, acceleration and course angle of the vehicle are calculated respectively. The method for compressing the track information comprises the following steps: for multi-frame track information corresponding to each second in the video, different vehicles and pedestrians are distinguished according to track IDs, and only track information of each vehicle and pedestrian when the vehicle and the pedestrian occur for the first time is reserved.
In this embodiment, the frame rate of the video is 25 frames per second, and the method for compressing the track information is as follows: every time continuous 25 frames of data are accumulated, different vehicles are distinguished according to track IDs, and only track information of each vehicle when the vehicle appears for the first time is reserved.
Specifically, the track information includes: center point pixel coordinates of the vehicle in the image, detection frame pixel height, detection frame pixel width, vehicle track ID, and frame number. Wherein the frame number is the frame number of the vehicle when first occurring in consecutive 25 frames.
In this embodiment, the method for calculating the speed, acceleration, and heading angle is as follows: different vehicles are distinguished by track ID, and the same vehicle is arranged at two adjacent moments t 1 、t 2 (t 1 ＜t 2 ) The two pieces of track information in (a) are respectively: t is t 1 Moment center point pixel abscissa x 1 Pixel ordinate y 1 Detecting the frame pixel height h 1 Detection frame pixel width w 1 ，t 2 Moment center point pixel abscissa x 2 Pixel ordinate y 2 Detecting the frame pixel height h 2 Detection frame pixel width w 2 。
Since there is a size difference between the far vehicle and the near vehicle in the image, the scale normalization coefficient is first calculated according to the following formula:
the speed is calculated according to the following formula:
wherein v is t2 x At t 2 Normalized component speed, v of time x-axis direction t2 y At t 2 Normalized component speed, v of moment y-axis direction t2 At t 2 Time of day speed.
The acceleration is calculated according to the following formula:
the heading angle is calculated as follows:
where σ is a very small positive number.
In this embodiment, different vehicles are distinguished by track ID, and a data structure is maintained to record the latest vehicle status of each vehicle, where the vehicle status includes: frame number, pixel abscissa, pixel ordinate, detection frame pixel height, detection frame pixel width, speed, acceleration, and heading angle. For all vehicles at each moment, searching the vehicle state of the vehicle in the data structure according to track ID, and if the vehicle state does not exist, creating a new vehicle state; if the vehicle state exists, calculating the normalized motion state of the current moment according to the existing vehicle state and the current moment track information, and then updating the vehicle state.
In particular, the data structure is implemented using dictionary nesting in python. When track information of a vehicle appears for the first time, recording the pixel abscissa, the pixel ordinate, the pixel height of the detection frame and the pixel width of the detection frame; when track information of a vehicle appears for the second time, calculating a size normalization coefficient according to a formula (1) according to the existing vehicle state and the track information of the current moment, calculating speeds according to a formula (2), a formula (3) and a formula (4) in sequence, and then updating the vehicle state; when track information of a vehicle appears for the third time, calculating a size normalization coefficient according to the existing vehicle state and the track information of the current moment, calculating the speed according to the formula (2), the formula (3) and the formula (4) in sequence, calculating the acceleration according to the formula (5), calculating the course angle according to the formula (6), and updating the vehicle state. When the track information of a vehicle appears three times, all variables in the vehicle state are assigned, and when the track information of the vehicle appears again, the vehicle state is updated in the above manner.
Step 3: and screening abnormal vehicles according to different motion indexes by using an isolated forest algorithm. And sending the speeds, the accelerations and the heading angles of all the vehicles at the same moment into three isolated forest models respectively to obtain vehicles with abnormal speeds, accelerations and heading angles. The three isolated forest models are obtained through training in advance; in this embodiment, the training method of the isolated forest model is as follows:
step (1): acquiring vehicle position information in each frame of the video by using a YOLOX target detection algorithm;
step (2): performing data association on the vehicle positions in different frames by using a ByteTrack multi-target tracking algorithm to obtain the motion trail of each vehicle and pedestrian in the video;
step (3): calculating the motion indexes of each vehicle according to the motion trail of each vehicle, wherein the motion indexes comprise speed, acceleration and course angle;
step (4): three isolated forest models are initialized by using the Isolationforest class in the open source machine learning toolkit scikit-learn, and then the isolated forest models are trained according to the speed, the acceleration and the course angle respectively by using the fit method of the Isolationforest class. The three obtained isolated forest models can be used for screening abnormal speeds, abnormal accelerations and abnormal course angles respectively.
In this embodiment, the method for screening abnormal vehicles according to different motion indexes by using an isolated forest algorithm is that the speed, the acceleration and the heading angle of each vehicle at the same moment are respectively sent into a corresponding isolated forest model, and the abnormal speed, the abnormal acceleration and the abnormal heading angle are respectively obtained by using an isolation_function method of the isolation forest class. Specifically, for a set of input speeds, accelerations or heading angles, the resolution_function method of the IsolationForest class correspondingly outputs a set of score efficiency numbers, the speed, acceleration or heading angle corresponding to the score efficiency number smaller than 0 is judged to be an abnormal value, and the corresponding vehicle is marked as speed abnormality, acceleration abnormality or heading angle abnormality.
Step 4: according to the screening result of the isolated forest algorithm, a scoring mechanism is provided for accumulating abnormal scores for each abnormal vehicle. In this embodiment, the scoring mechanism proposed is: according to the screening result of the isolated forest algorithm, calculating the abnormal score of the motion index of the vehicle with abnormal motion index, judging whether the vehicle environment is abnormal, if so, calculating the abnormal score of the vehicle environment, and finally averaging the abnormal score and the vehicle environment to obtain the abnormal score, wherein the method specifically comprises the following steps: step (1): for any moment, calculating the abnormal score of the motion index according to the following formula according to the vehicle with abnormal motion index output by the isolated forest algorithm at the current moment:
S motion ＝α·S speed +β·S acc +γ·S θ’ (α+β+γ＝1), (7)；
wherein alpha, beta and gamma are respectively abnormal speed weight, abnormal acceleration weight and abnormal heading angle weight, and the values are respectively 0.4, 0.2 and 0.4; s is S speed 、S acc 、S θ Respectively, a speed anomaly score, an acceleration anomaly score and a course angle anomaly score, and S is the time of abnormal vehicle speed speed 1, otherwise 0; s when the acceleration of the vehicle is abnormal acc 1, otherwise 0; s when the heading angle of the vehicle is abnormal θ 1, otherwise 0.
Step (2): for each abnormal vehicle, calculating the normalized distance between the abnormal vehicle and all pedestrians at the moment according to the following formula, and selecting the minimum value, wherein the minimum value is the pedestrian closest to the normalized distance of the abnormal vehicle, and if the minimum value is smaller than 1, judging that the vehicle environment of the vehicle is abnormal;
x in the above person For the centre point pixel abscissa, y of a person person The ordinate of the pixel of the center point of the person, h person Human detection frame pixel height, w person Human detection frame pixel width, x car For the centre point pixel abscissa, y of the vehicle car Is the ordinate of the pixel of the central point of the vehicle, h car Detecting frame pixel height, w for vehicle car Is the detection frame pixel width of the vehicle.
Step (3): for each abnormal vehicle, if the vehicle environment is not abnormal, namely the minimum value in the normalized distance between all pedestrians and the vehicles is greater than 1, the vehicle environment abnormality is divided into 0, otherwise, the vehicle environment abnormality is calculated according to the following formula:
S distance ＝1-d， (9)；
step (4): the anomaly score for the vehicle is calculated according to the following formula:
according to the above calculation method, the maximum value of the abnormality score that can be obtained by each abnormal vehicle at each time is 1.
The abnormal vehicle is determined and scored only when the vehicle motion index is abnormal, and the vehicle having an abnormal vehicle environment but a normal motion index is not determined as an abnormal vehicle.
Step 5: an attention reassignment mechanism is proposed, with which the resulting attention coefficients weight the outliers of each vehicle to highlight the accident vehicles among all outlier vehicles at the same time. As shown in the schematic diagram of the attention re-allocation mechanism in fig. 3, the attention re-allocation mechanism is as follows: naturally, the attention coefficient of each abnormal vehicle before attention transfer is 1, and the total attention value at each moment is equal to the number of abnormal vehicles at the moment; for each abnormal vehicle, searching the abnormal times in the early warning ID pool according to the track ID, and calculating the weight W according to the abnormal times t The method comprises the steps of carrying out a first treatment on the surface of the For each abnormal vehicle, dividing the abnormal score by the maximum value of all vehicle abnormal scores at the current moment to obtain a normalized abnormal score, and calculating a weight W according to the normalized abnormal score s The method comprises the steps of carrying out a first treatment on the surface of the For each abnormal vehicle, pairW of which is t And weight W s Obtaining the comprehensive weight of the geometric mean value; converting the comprehensive weights of all vehicles into attention redistribution coefficients using a softmax function; for each abnormal vehicle, multiplying the attention redistribution coefficient by the total attention value to finish attention redistribution, wherein the attention redistribution specifically comprises:
step (1): maintaining an early warning ID pool to record the early warning times of each vehicle, for each abnormal vehicle at each moment, inquiring the abnormal times T in the early warning ID pool according to track ID thereof, and mapping the abnormal times T into weights W according to the following formula t ：
Step (2): for any moment, the maximum abnormal score S of all abnormal vehicles at the current moment is counted max For each abnormal vehicle at the current moment, mapping its abnormal score S to weight W according to the following s ：
In the above, the abnormality score S of each vehicle is divided by S max Normalize its outlier to (0, 1)]Mapping the normalized anomaly score to a weight W using a sinusoidal function s ；
Step (3): for each abnormal vehicle, setting the total number of abnormal vehicles at the current moment as N, and calculating the attention coefficient of each abnormal vehicle after the attention is redistributed according to the following formula:
in the above, for each abnormal vehicle, weight W is given to t And weight W s And obtaining the comprehensive weight by taking the geometric average value. Converting the integrated weights of all vehicles into attention redistribution coefficients using softmax function, then paying attention to that momentAnd multiplying the total force value by the attention redistribution coefficient of each abnormal vehicle to obtain the attention coefficient of the corresponding abnormal vehicle after attention redistribution.
Referring to fig. 4 and 5, fig. 4 is a vehicle anomaly score line graph when the attention reassignment mechanism is not used, and fig. 5 is a vehicle anomaly score line graph when the attention reassignment mechanism is used, in which the horizontal axis is the frame number, the vertical axis is the anomaly score, and the tag is the track ID of the anomaly vehicle. It can be seen that, compared with the case that the attention redistribution mechanism is not used, the abnormal gap between the accident vehicle and other abnormal vehicles in the abnormal vehicles can be effectively increased by using the mechanism, and the abnormal vehicles can be more favorably distinguished after the score efficiency is calculated.
Step 6: the video is divided into a plurality of judging periods, in each judging period, the score efficiency of each abnormal vehicle is calculated according to the accumulated abnormal score, the abnormal vehicle with the score efficiency exceeding the alarm threshold value is judged as an accident vehicle, and meanwhile, the early warning times of each abnormal vehicle are accumulated according to whether the score efficiency of each abnormal vehicle exceeds the early warning threshold value or not according to track ID to be used as the calculation basis of the attention transfer mechanism. In the present embodiment, the score efficiency of each abnormal vehicle is calculated from the accumulated abnormal score every 250 frames.
Specifically, in the previous step two, mention has been made of: the frame rate of the video is 25 frames per second, track information of every 25 frames is compressed into track information of 1 moment, and the number of moments of 250 continuous frames is 10. For each abnormal vehicle, the product of the attention coefficient and the abnormal score at all times in the continuous frame is accumulated, and then divided by the time number of the continuous frame to obtain the score efficiency of the abnormal vehicle as the accident vehicle in the continuous frame. The calculation formula is as follows:
in the above, L is the total time number of the continuous frames, W i j And S is i j Attention coefficient and abnormality score of the ith abnormal vehicle at the jth moment respectively, C i I.e. the ith abnormal vehicle is in the 250 consecutive framesInside is the scoring efficiency of the accident vehicle.
In this embodiment, for each abnormal vehicle, if the score efficiency is greater than 0.5, the early warning times are accumulated in the early warning ID pool as the calculation basis of the attention redistribution mechanism at the subsequent time; if the score efficiency is greater than 1, it is determined as an accident vehicle.
Referring to fig. 6, in the embodiment, the detection result is visualized by using an open source computer vision library opencv. Specifically, track information of the vehicle is stored in a txt file, and file contents include: frame number, vehicle track ID, detection frame position; storing the detection result in a json file, wherein the file content comprises: judging the length of the frame segment, the starting frame number of the frame segment, track ID of each abnormal vehicle of the frame segment and scoring efficiency; and reading and combining the detection results in the vehicle track information in the txt file and the json file, marking track IDs and detection frames of all vehicles in the original video by using a rectangle function in an opencv library, and filling the detection frames of the accident vehicles with red.
The evaluation indexes used in this embodiment include: classification accuracy (CR), detection Rate (DR), specificity, false Alarm Rate (FAR), precision (Precision), and F-metric (F1-Score). The above indices need to be calculated with the aid of a confusion matrix, which is shown in table 1. Wherein True Positive (TP) indicates the number of Positive class events correctly classified, true Negative (TN) indicates the number of Negative class events correctly classified, false Positive (FP) indicates the number of Positive class events incorrectly classified, and False Negative (FN) indicates the number of Negative class events incorrectly classified.
TABLE 1 confusion matrix
In this embodiment, for each frame in the video, if the algorithm alarms, the Prediction classification (Prediction Class) is Positive (Positive), otherwise the Prediction classification (Prediction Class) is Negative (Negative), if there is a traffic accident, the Actual classification (Actual Class) is correct (True), otherwise the Actual classification (Actual Class) is incorrect (False). On this basis, the meaning and calculation method of each evaluation index are as follows:
(1) Classification accuracy (Classification Rate, CR for short)
The classification accuracy is the percentage of the number of correctly classified events to the number of all the occurrence events, and the calculation formula is as follows:
(2) Detection Rate (DR for short)
The detection rate is the percentage of the number of correctly detected events to the number of practically all the occurring events, and the calculation formula is as follows:
(3) Specificity (Specificity)
The specificity is the percentage of the number of correctly detected negative events to the number of negative events actually happening, and the calculation formula is as follows:
(4) False alarm Rate (FAR for short)
The false alarm rate is the percentage of the number of the false predicted correct events to the number of the negative events actually happening, and the calculation formula is as follows:
(5) Precision (Precision)
The precision is the percentage of the number of detected positive events to the number of the positive events actually happening, and the calculation formula is as follows:
(6) F metric (F1-Score)
F, measuring a harmonic mean value of the precision and the detection rate, wherein a calculation formula is as follows:
in this embodiment, the algorithm was tested using a self-built dataset with and without the attention re-allocation mechanism, respectively, and the experimental results are shown in tables 2, 3.
TABLE 2 traffic accident detection index employing attention redistribution mechanism
TABLE 3 traffic accident detection index without attention redistribution mechanism
As can be seen from tables 2 and 3, as the alarm threshold increases, the classification accuracy (CR), the Detection Rate (DR), the False Alarm Rate (FAR), and the F metric (F1-Score) gradually decrease, and the Specificity and the accuracy (Precision) gradually increase. It should be noted that, because the adoption of the attention redistribution mechanism can lead to the improvement of the upper score efficiency, the alarm threshold ranges selected by the two experiments are slightly different. Comparing the two tables, it can be found that, compared with adopting the attention redistribution mechanism, six indexes are difficult to reach better level at the same time without adopting the mechanism, because the difference of the accident vehicle and other vehicles in the scoring efficiency is not big without adopting the mechanism, therefore, when a higher alarm threshold is selected, a large number of accident vehicles are screened out while normal vehicles are screened out, so that although the indexes of Specificity, false Alarm Rate (FAR) and Precision (Precision) are better, the indexes of classification accuracy (CR), detection Rate (DR) and F measurement (F1-Score) are poor; in contrast, when a lower alarm threshold is selected, a large number of normal vehicles are reserved while most accident vehicles are reserved, so that indexes of classification accuracy (CR), detection Rate (DR) and F measurement (F1-Score) are good, but indexes of Specificity, false Alarm Rate (FAR) and Precision (Precision) are poor. In contrast, when the attention reassignment mechanism is adopted, each index is stable and performs better under different alarm thresholds.
In addition, in the present embodiment, the frame rate (Frames Per Second, FPS for short) of the algorithm is about 27.53 frames per second, which is greater than the video frame rate of 25 frames per second, so that the real-time detection requirement can be satisfied.
The foregoing is a further detailed description of the invention in connection with the preferred embodiments, and it is not intended that the invention be limited to the specific embodiments described. It will be apparent to those skilled in the art that several simple deductions or substitutions may be made without departing from the spirit of the invention, and these should be considered to be within the scope of the invention.

Claims (7)

1. A traffic accident detection method based on an isolated forest algorithm and target tracking is characterized in that: comprises the following steps of the method,
step 1: for input traffic video data, firstly, acquiring the positions of vehicles and pedestrians in each frame of image by using a target tracking technology, and carrying out data association on the positions of the vehicles and pedestrians in different frames to obtain the motion trail of each vehicle and pedestrian in the video;
step 2: calculating the motion index of each vehicle according to the motion trail of each vehicle;
step 3: screening abnormal vehicles according to different motion indexes by using an isolated forest algorithm;
step 4: according to the screening result of the isolated forest algorithm, a scoring mechanism is provided for accumulating abnormal scores of each abnormal vehicle;
step 5: the attention redistribution mechanism is provided, the obtained attention coefficient weights the abnormal parts of all vehicles, and accident vehicles in all abnormal vehicles at the same moment are highlighted;
step 6: the video is divided into a plurality of discrimination periods, and in each discrimination period, the score efficiency of each abnormal vehicle is calculated from the accumulated abnormal score, and the accident vehicle is determined from the score efficiency.
2. The traffic accident detection method based on the isolated forest algorithm and the target tracking according to claim 1, wherein: in the step 1, traffic accident detection is performed based on single-mode data of the video, and a target tracking algorithm is adopted to obtain motion tracks of vehicles and pedestrians in the video data, wherein the motion tracks of each vehicle and pedestrian in each frame of image are in the following specific forms: center point pixel coordinates, detection frame pixel height, detection frame pixel width, track ID, and frame number.
3. The traffic accident detection method based on the isolated forest algorithm and the target tracking according to claim 1, wherein in the step 2, the motion indexes of the vehicles are calculated according to the motion track of each vehicle, firstly, the motion track information of the vehicles and the pedestrians is compressed in the time dimension, for the multi-frame track information corresponding to each second in the video, different vehicles and pedestrians are distinguished according to track IDs, only the track information of each vehicle and each pedestrian when the vehicle and the pedestrians occur for the first time is reserved, the minimum time unit after track compression is called as time, and then the speed, the acceleration and the course angle of the vehicle are calculated according to the positions of each vehicle at a plurality of time;
the method comprises the following specific steps: different vehicles are distinguished by track ID, and the same vehicle is arranged at two adjacent moments t 1 、t 2 (t 1 ＜t 2 ) The two pieces of track information in (a) are respectively: t is t 1 Moment center point pixel abscissa x 1 Pixel ordinate y 1 Detecting the frame pixel height h 1 Detection frame pixel width w 1 ；t 2 Moment center point pixel abscissa x 2 Pixel ordinate y 2 Detecting the frame pixel height h 2 And a detection frame pixel width w 2 ，
The scale normalization coefficient is calculated as follows:
the calculation speed is as follows:
wherein v is t2 x At t 2 Normalized component speed, v of time x-axis direction t2 y At t 2 Normalized component speed, v of moment y-axis direction t2 At t 2 Time speed;
the acceleration is calculated as follows:
the calculated heading angle is as follows:
where σ is a very small positive number.
4. The traffic accident detection method based on the isolated forest algorithm and the target tracking according to claim 1, wherein in the step 3, abnormal vehicles are screened according to different motion indexes by using the trained isolated forest algorithm, and the speed, the acceleration and the heading angle of all vehicles at the same moment are respectively sent into three isolated forest models to obtain the vehicles with abnormal speeds, accelerations and heading angles.
5. The traffic accident detection method based on the isolated forest algorithm and the target tracking according to claim 1, wherein in the step 4, a scoring mechanism is provided for accumulating abnormal scores for each abnormal vehicle according to the screening result of the isolated forest algorithm, and the scoring mechanism comprises: calculating motion index anomaly score and vehicle environment anomaly score for vehicles with motion indexes abnormal, judging whether the vehicle environment is abnormal, calculating vehicle environment anomaly score if abnormal, and averaging the motion index anomaly score and the vehicle environment anomaly score to obtain anomaly score;
the motion index anomaly score calculation formula is as follows:
S motion ＝α·S speed +β·S acc +γ·S θ ，(α+β+γ＝1),
in the above, alpha, beta and gamma are respectively abnormal speed weight, abnormal acceleration weight and abnormal course angle weight; s is S speed 、S acc 、S θ Respectively, a speed anomaly score, an acceleration anomaly score and a course angle anomaly score, and S is the time of abnormal vehicle speed speed 1, otherwise 0; s when the acceleration of the vehicle is abnormal acc 1, otherwise 0; s when the heading angle of the vehicle is abnormal θ 1, otherwise 0;
the vehicle environment abnormality judging mode is as follows: when people exist nearby the vehicle and the normalized distance of the people and the vehicle is smaller than 1, scoring is carried out, and the calculation formula of the normalized distance of the people and the vehicle is as follows:
x in the above person For the centre point pixel abscissa, y of a person person The ordinate of the pixel of the center point of the person, h person Human detection frame pixel height, w person Human detection frame pixel width, x car For the centre point pixel abscissa, y of the vehicle car Is the ordinate of the pixel of the central point of the vehicle, h car Detecting frame pixel height, w for vehicle car The width of the detection frame pixels of the vehicle;
for each abnormal vehicle, if the vehicle environment is not abnormal, namely the minimum value in the normalized distance between all pedestrians and the vehicles is greater than 1, the vehicle environment abnormality is divided into 0, otherwise, the vehicle environment abnormality is calculated according to the following formula:
S distance ＝1-d min ，
in the above, d min Normalizing the minimum value in the distances between all pedestrians and the vehicles;
the calculation formula of the abnormal score of the vehicle is as follows:
in the above, S motion S is the abnormal division of movement index distance Is an abnormal component of the vehicle environment.
6. The traffic accident detection method based on the isolated forest algorithm and the target tracking according to claim 1, wherein the attention reassignment mechanism in step 5 transfers the attention coefficient part of the abnormal vehicles at the same time to other abnormal vehicles, weights the abnormal parts of each vehicle by using the obtained attention coefficient, and highlights the accident vehicles in all the abnormal vehicles at the same time; wherein the attention redistribution mechanism is: the attention coefficient of each abnormal vehicle before the attention transfer is 1, and the total attention value of each moment is equal to the number of abnormal vehicles at the moment; for each abnormal vehicle, inquiring the early warning times T before the current moment, and early warning the abnormal vehicle according to the following formulaThe times T are mapped into weights W t ：
Statistics of maximum anomaly score S of all anomaly vehicles at current moment max For each abnormal vehicle at the current moment, mapping its abnormal score S to weight W according to the following s ：
For each abnormal vehicle, setting the total number of abnormal vehicles at the current moment as N, and calculating the attention coefficient of each abnormal vehicle after the attention is redistributed according to the following formula:
7. the traffic accident detection method based on the isolated forest algorithm and the target tracking according to claim 1, wherein in the step 6, the video is divided into a plurality of discrimination periods, the score efficiency of each abnormal vehicle is calculated according to the accumulated abnormal score in each discrimination period, the accident vehicle is judged according to the score efficiency, wherein different abnormal vehicles are distinguished by track ID, the early warning times of each abnormal vehicle is accumulated according to whether the score efficiency of each abnormal vehicle exceeds the early warning threshold, the abnormal vehicle with the score efficiency exceeding the alarm threshold is judged as the accident vehicle, and the calculation formula of the score efficiency is as follows:
wherein L is the total time number of the continuous frames, W i j And S is i j The ith abnormal vehicle is at the jth timeAttention coefficients and anomaly scores of the score.
CN202310596197.9A 2023-05-24 2023-05-24 Traffic accident detection method based on isolated forest algorithm and target tracking Pending CN116681722A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310596197.9A CN116681722A (en) 2023-05-24 2023-05-24 Traffic accident detection method based on isolated forest algorithm and target tracking

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310596197.9A CN116681722A (en) 2023-05-24 2023-05-24 Traffic accident detection method based on isolated forest algorithm and target tracking

Publications (1)

Publication Number Publication Date
CN116681722A true CN116681722A (en) 2023-09-01

Family

ID=87777969

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310596197.9A Pending CN116681722A (en) 2023-05-24 2023-05-24 Traffic accident detection method based on isolated forest algorithm and target tracking

Country Status (1)

Country Link
CN (1) CN116681722A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117012032A (en) * 2023-09-28 2023-11-07 深圳市新城市规划建筑设计股份有限公司 Intelligent traffic management system and method based on big data

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117012032A (en) * 2023-09-28 2023-11-07 深圳市新城市规划建筑设计股份有限公司 Intelligent traffic management system and method based on big data
CN117012032B (en) * 2023-09-28 2023-12-19 深圳市新城市规划建筑设计股份有限公司 Intelligent traffic management system and method based on big data

Similar Documents

Publication Publication Date Title
CN109829403B (en) Vehicle anti-collision early warning method and system based on deep learning
Aboah A vision-based system for traffic anomaly detection using deep learning and decision trees
CN110738127B (en) Helmet identification method based on unsupervised deep learning neural network algorithm
CN112700470B (en) Target detection and track extraction method based on traffic video stream
JP6570731B2 (en) Method and system for calculating passenger congestion
CN109948582B (en) Intelligent vehicle reverse running detection method based on tracking trajectory analysis
CN110070074B (en) Method for constructing pedestrian detection model
CN111898501B (en) Unmanned aerial vehicle online aerial photography vehicle identification and statistics method for congested road sections
CN111242015B (en) Method for predicting driving dangerous scene based on motion profile semantic graph
CN116681722A (en) Traffic accident detection method based on isolated forest algorithm and target tracking
CN110633643A (en) Abnormal behavior detection method and system for smart community
CN110781873A (en) Driver fatigue grade identification method based on bimodal feature fusion
CN112084928A (en) Road traffic accident detection method based on visual attention mechanism and ConvLSTM network
CN114170580A (en) Highway-oriented abnormal event detection method
CN112435276A (en) Vehicle tracking method and device, intelligent terminal and storage medium
CN113034378A (en) Method for distinguishing electric automobile from fuel automobile
CN113313012B (en) Dangerous driving behavior identification method based on convolution generation countermeasure network
CN112700473A (en) Carriage congestion degree judging system based on image recognition
CN116109986A (en) Vehicle track extraction method based on laser radar and video technology complementation
CN114822044B (en) Driving safety early warning method and device based on tunnel
CN115907462A (en) Intersection safety state sensing and diagnosis treatment system and method based on unmanned aerial vehicle video
CN115601558A (en) Single turnout state detection system and detection method and semi-automatic data labeling method
KR20220067271A (en) Image acquisition apparatus and image acquisition method
CN114550088B (en) Multi-camera fused passenger identification method and system and electronic equipment
Xun et al. A method of vehicle trajectory tracking and prediction based on traffic video

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