CN109948582B - Intelligent vehicle reverse running detection method based on tracking trajectory analysis - Google Patents
Intelligent vehicle reverse running detection method based on tracking trajectory analysis Download PDFInfo
- Publication number
- CN109948582B CN109948582B CN201910244210.8A CN201910244210A CN109948582B CN 109948582 B CN109948582 B CN 109948582B CN 201910244210 A CN201910244210 A CN 201910244210A CN 109948582 B CN109948582 B CN 109948582B
- Authority
- CN
- China
- Prior art keywords
- target vehicle
- vehicle
- target
- tracking
- classifier
- 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.)
- Active
Links
Images
Abstract
The invention discloses a vehicle retrograde motion intelligent detection method based on tracking track analysis, which comprises the steps of detecting target vehicles in a video image by utilizing a primary classifier and a secondary classifier, extracting a target vehicle region, distributing nuclear-related filtering trackers for the target vehicles, matching each target vehicle with one nuclear-related filtering tracker, obtaining a tracking region of the target vehicle by utilizing the nuclear-related filtering trackers, further obtaining a motion track growth direction of the target vehicle according to the tracking region, comparing the motion track growth direction with the retrograde motion direction of an initial mark, and if the motion track growth direction is the same as the motion track growth direction, enabling the target vehicles to carry out retrograde motion; if not, the target vehicle does not run backwards. The invention realizes the real-time detection of the vehicle retrograde motion by the mode, and solves the problems of manual detection; meanwhile, the reliability of vehicle detection and identification is improved by using the cascade classifier, and the simultaneous tracking of multiple vehicle targets is realized.
Description
Technical Field
The invention belongs to the technical field of computer vision, and particularly relates to an intelligent vehicle retrograde motion detection method based on tracking trajectory analysis.
Background
With the continuous development of society, vehicles on roads are more and more, the traffic safety problem is more and more prominent, and traffic accidents caused by retrograde motion on the roads often cause great harm each year because of the surprise of casualties of the traffic accidents. The attention on road safety is continuously increased at present, a large number of cameras are installed on roads for monitoring, and whether abnormity occurs or not is mostly judged by manpower for the massive monitoring video data, so that a large amount of manpower is consumed, and uninterrupted work can not be performed within 24 hours. With the development of intelligent video monitoring technology, some basic road traffic abnormal events can be automatically detected.
At present, most of existing vehicle retrograde motion detection methods extract a moving target by using methods such as an optical flow method, background modeling and the like, then realize vehicle tracking through similarity matching, and further judge whether the vehicle is retrograde by using vehicle track information. This method is simple to implement, but for the optical flow method, if the speed of the moving object is slow to detect, it is difficult, and the background modeling is more susceptible to the environment. In the current research method for detecting and tracking the vehicle, the prior knowledge is used for detecting the vehicle by utilizing the prior knowledge such as shadow at the bottom of the vehicle, tail lamp and the like aiming at the traditional method of a detection part; or training classifiers by using a higher-level Haar _ like feature and an HOG feature to detect the vehicle, but a single classifier trained by using a single feature is difficult to adapt to a complex and variable traffic scene; or, the current popular detection method based on the convolutional neural network is used, although the effect is obviously improved, the real-time operation on a common hardware equipment terminal is difficult due to large calculation amount. For tracking of vehicle targets, most of the conventional methods use kalman filtering, particle filtering and the like, and the methods generally can only track a single target, and once a tracking error occurs, the method cannot correct the tracking error in time, so that the effect is unreliable.
Therefore, an efficient and reliable vehicle retrograde motion detection method is lacked in the prior art, the classification reliability of a single classifier in the prior vehicle detection technology needs to be improved, and the prior vehicle tracking method is a defect in tracking a single target, so that a vehicle retrograde motion detection method based on the reliable vehicle detection and tracking technology is urgently needed.
Disclosure of Invention
The invention provides an intelligent detection method for vehicle retrograde motion based on tracking trajectory analysis, which can realize the rapid and accurate identification of vehicles retrograde motion on a road, and simultaneously, in the vehicle detection process, two off-line trained classifiers are used for secondary identification, so that the reliability of the vehicle detection result is improved; meanwhile, the tracking of multiple vehicle targets is realized, and the application range of the system is expanded.
On one hand, the invention provides an intelligent detection method for vehicle reverse running based on tracking track analysis, which comprises the following steps:
s1: acquiring a retrograde motion detection area and a retrograde motion direction mark on a shooting area image of a camera and acquiring video data of the camera in real time;
the retrograde motion detection area is a quadrilateral area containing a vehicle driving lane, and the retrograde motion direction is marked as a directional line segment;
s2: extracting a target vehicle region in a reverse detection region from a current frame of image of video data based on a first-stage classifier and a second-stage classifier of off-line training;
the method comprises the steps that a first-stage classifier is used for classifying and identifying a current frame image to obtain a target vehicle candidate region, and a second-stage classifier is used for classifying and identifying the target vehicle candidate region to obtain a target vehicle region on the current frame image;
the input data of the first-stage classifier and the second-stage classifier in the off-line training process are respectively a positive sample image and a negative sample image, wherein the positive sample image contains vehicles, and the negative sample image does not contain vehicles;
s3: based on the target vehicle region extracted in step S2, distributing relevant filtering trackers, each kernel relevant filtering tracker tracking the target vehicle based on the target vehicle region in the current frame image and other frame images in the video data to obtain tracking information of the corresponding target vehicle;
a core related filtering tracker tracks a target vehicle, and all current core related filtering trackers and corresponding target vehicles form a current tracking queue;
sequentially processing other frame images in the video data according to the step S2 and the step S3, wherein if an identified target vehicle is not in the current tracking queue in the processing process, the target vehicle is a new target vehicle, and a kernel-related filtering tracker is distributed for tracking the new target vehicle;
s4: identifying whether the target vehicle has retrograde motion in the retrograde motion detection area based on the tracking information of the target vehicle and the retrograde motion direction mark;
the tracking information comprises target detection frame positions of the target vehicle in the continuous frame images, and the target detection frames represent tracking areas of the target vehicle in each frame image; acquiring the growth direction of a motion trail at the same position on a target detection frame based on the position of the target detection frame of a target vehicle in a continuous frame image, judging whether the growth direction of the motion trail is the same as the retrograde direction, and if so, performing retrograde motion corresponding to the target vehicle; if not, no retrograde motion occurs.
The invention tracks the motion track growth direction of the target vehicle and compares the motion track growth direction with the initially set retrograde direction mark to further draw the conclusion whether the vehicle is retrograde, thereby providing a means for automatically identifying whether the vehicle is retrograde; meanwhile, the target vehicle is identified based on the classifiers of the two levels, and compared with the identification result of a single classifier, the reliability of the identification result obtained by the method provided by the invention is higher. In addition, the mode provided by the invention realizes multi-target tracking in the tracking process, and each vehicle target is allocated with one tracker which is added and matched with the actual application.
Further preferably, if the vehicle does not run backwards, the method further comprises identifying the running state of the target vehicle based on the tracking information of the target vehicle, and the process is as follows:
firstly, acquiring the motion trail growth length of the same position on a target detection frame based on the position of the target detection frame of a target vehicle in continuous N1 frame images; then, judging whether the growth length of the motion trail is greater than a preset first threshold value L1;
if the current speed is greater than a preset first threshold value L1, the target vehicle runs normally;
otherwise, acquiring the growth length of a new motion track at the same position on the target detection frame based on the position of the target detection frame of the target vehicle in the continuous N2 frame images, and judging whether the growth length of the new motion track is smaller than a preset second threshold value L2;
wherein N2 is greater than N1, and L2 is less than L1;
if the current time interval is smaller than a preset second threshold value L2, judging whether other target vehicles exist in the same driving direction at the current same time interval;
if other target vehicles exist, acquiring the motion track growth lengths of the other target vehicles in the same driving direction in the same time period in the continuous N1 frames;
if the motion track growth lengths of other target vehicles in the continuous N1 frames are greater than a preset first threshold value L1, the target vehicles are in an abnormal state; if the motion trail growth lengths of other target vehicles in the continuous N1 frames are all smaller than a preset second threshold value L2, the target vehicles are in a blocked state; if the motion trail growth lengths of other targets in the continuous N1 frames are less than or equal to a preset first threshold L1 and greater than or equal to a second threshold L2, the target vehicle is in a pending state;
if no other target vehicle exists, the target vehicle is in an abnormal state;
and if the growth length of the new motion trail is not less than a preset second threshold value L2, the target vehicle is in a pending state.
The invention can also identify the state of the target vehicle by utilizing the growth length of the motion trail, and timely check out abnormal conditions so as to be convenient for timely regulation and control, thereby providing a certain basis for relieving or solving traffic problems.
Further preferably, the value ranges of the preset first threshold L1 and the preset second threshold L2 corresponding to the target vehicle are respectively as follows: l1>2/3h, L2<1/3 h;
and h is the high pixel value of the target detection frame when the tracker captures the target vehicle for the first time.
As can be seen from the above description, the preset first threshold L1 and the preset second threshold L2 are determined according to the target detection frames of the target vehicles, that is, the first threshold and the second threshold of the motion trail length of each tracked target vehicle are different, because the sizes of the target detection frames are different due to different vehicle types. However, the values of the preset first threshold L1 and the preset second threshold L2 are consistent.
The values of the N1 frames and the N2 frames are empirical values, and are related to the actual video acquisition frame rate, for example, the larger the acquisition frame rate of the camera is, the larger the values of the N1 frames and the N2 frames are. Generally, the effective detection distance of a monitoring camera is about 100 meters, the vehicle appearance time is about 3 seconds according to a high-speed vehicle speed of 120km/h, the algorithm is supposed to effectively identify that the vehicle time is only 2s, the video acquisition frame rate is m fps (m frames per second), the value of N1 can be selected to be m, and the value of N2 is 2 m.
Further preferably, in step S4, the motion trajectory growth direction and length of the lower center point position of the target detection frame are obtained based on the position of the target detection frame of the target vehicle in the continuous frame images;
wherein, every interval n1 frames records the coordinate value of the lower center point position of the target detection frame.
And recording coordinate values of the lower center point position of the target detection frame at every n1 frames, and further connecting in series to obtain the growth direction and the length of the motion trail. The coordinate values are recorded at intervals of n1 frames instead of the coordinate values corresponding to each frame of image, so that data redundancy can be placed, space waste is reduced, and the representation of the target motion track cannot be lost. n1 is a positive integer greater than or equal to 2, such as n1 is 5.
Further preferably, the method further includes the following steps when processing the next frame of image:
firstly, masking a tracking area where a target vehicle in a current tracking queue in a next frame of image is located;
acquiring a tracking area of a target vehicle in a next frame of image by using a nuclear correlation filtering tracker of the target vehicle in a current tracking queue;
then, the masked next frame image is processed according to step S2 and step S3.
The masking method in the invention is to set the pixel values of the pixel points corresponding to the tracking areas to be 255. Therefore, when the target vehicle region detection is carried out on the next frame of image, the identified target vehicle is the new target vehicle. The corresponding tracking area can be judged as a negative sample (namely a non-target vehicle area) in the primary classifier and the secondary classifier by a mask mode, and the detection speed is increased.
Further preferably, when the tracking information of the corresponding target vehicle is acquired by using the nuclear correlation filtering tracker, whether the target vehicle disappears in the visual field or reaches the boundary position of the retrograde detection area is supervised;
if the vehicle disappears in the visual field range or the target vehicle reaches the boundary position of the reverse driving detection area, deleting the kernel-related filtering tracker in the current tracking queue;
wherein the tracking area of the target vehicle in the next frame can be predicted by the kernel correlation filtering tracker, and the visual field range and the position of the reverse detection area are determined, so that whether the target vehicle is in the visual field range or still in the reverse detection area can be distinguished.
Further preferably, when the tracking information of the corresponding target vehicle is obtained by using the nuclear-related filtering tracker, punishment scoring is carried out on tracking of the nuclear-related filtering tracker;
acquiring a tracking area of a target vehicle in a next frame of image based on a kernel correlation filtering tracker, and identifying whether the tracking area contains the vehicle by using a vehicle classifier trained offline in advance; if the vehicle is not included, the vision-check related filtering tracker makes a tracking error, and punishment scoring is carried out on the tracking error;
when the probability p is greater than or equal to 0.5, the vehicle is included, and when the probability p is less than 0.5, the vehicle is not included; when the vehicle classifier obtains that the probability of the vehicle is p, the score is 1-p;
if the accumulated score of the kernel-related filtering tracker reaches a preset S score, deleting the kernel-related filtering tracker in a current tracking queue;
the input data of the vehicle classifier training process are positive sample images and negative sample images.
The kernel correlation filter tracker will always have the object tracked (e.g. at frame n +30 the tracker has lost the correct object, but the tracker will continue to work, finding the closest to the area tracked by frame n +30 in frame n +31, although this is already an error). This is the target loss that often occurs in the tracking process of the kernel correlation filter tracker. The invention uses a vehicle classifier to identify whether a loss of a tracked target occurs. Meanwhile, the classifier cannot reach one hundred percent of accuracy in an actual complex traffic scene, so that the vehicle classifier judges whether the vehicle immediately tracks the vehicle and discards the vehicle from the tracker, but gives a penalty score, and the vehicle classifier discards the vehicle from the tracker when the penalty reaches a certain threshold S.
The vehicle classifier is a two-classifier, and the obtained result is a probability value p with the size of 0-1. The score is triggered only if the vehicle classifier deems it not to be a vehicle. In the present invention s is an empirical value, such as 2 in the present embodiment. In the invention, each nuclear-related filtering tracker only concerns one target vehicle, if a penalty threshold value S is reached, the nuclear-related filtering tracker is removed from a tracking queue, and a classifier and a score S corresponding to the nuclear-related filtering tracker are eliminated.
Further preferably, the first-stage classifier is an AdaBoost classifier based on a cascade model; the secondary classifier is a support vector machine;
training the training process of the primary classifier based on Haar _ like characteristics of positive samples and negative samples;
the training process of the secondary classifier is trained on the basis of feature vectors formed by HOG features of positive samples and negative samples.
Advantageous effects
On one hand, the method tracks the motion track growth direction of the target vehicle and compares the motion track growth direction with the initially set retrograde direction mark to further draw the conclusion whether the vehicle is retrograde, thereby providing a means for automatically identifying whether the vehicle is retrograde and getting rid of the obstacle of manual monitoring.
On the other hand, the invention can identify whether the vehicle is running in the wrong direction or not, and can also identify the current state of the vehicle to identify whether the vehicle is running normally or in an abnormal state or in a traffic jam state, so as to facilitate timely scheduling and provide a data basis for traffic management.
In addition, the process of reverse running identification and state identification is realized by the invention, the target vehicle is required to be accurately identified and tracked, and the target vehicle is identified by utilizing two grades of classifiers, so that the identification result with higher reliability can be obtained compared with the identification result of a single classifier. At the same time, the user can select the desired position,
the mode provided by the invention realizes multi-target tracking in the tracking process, and each vehicle target is allocated with one tracker which is in addition fit with the actual application.
Drawings
Fig. 1 is a schematic flow chart of an intelligent detection method for vehicle reverse running based on track analysis according to the present invention.
Fig. 2 is a schematic flow chart of vehicle detection tracking and trajectory analysis in the intelligent vehicle converse detection method based on tracking trajectory analysis.
Fig. 3 is a schematic diagram of the movement of the target detection frame in the vehicle movement track analysis according to the present invention.
Detailed Description
The present invention will be further described with reference to the following examples.
The invention provides a vehicle converse driving intelligent detection method based on tracking track analysis, which comprises the following steps:
s1: the method comprises the steps of obtaining a retrograde motion detection area and a retrograde motion direction mark on a shooting area image of a camera and obtaining video data of the camera in real time.
Since the lane position and the lane direction in the monitoring-installed image of the shot area are uniquely determined and fixed, the initialization process (manually marking the reverse detection area and the reverse direction) is only required to be performed once.
S2: and extracting the target vehicle area in the reverse detection area from the current frame of image of the video data based on the first-stage classifier and the second-stage classifier trained off line.
In the embodiment, a domestic vehicle detection data set is used, and in combination with collected traffic image data sets in local cities and expressways, vehicles are cut out to serve as positive samples, traffic scenes without vehicles are cut out to serve as negative samples, a positive sample set and a negative sample set used for training are manufactured, the proportion of the positive samples to the negative samples is about 1:3, and the positive samples and the negative samples are uniformly scaled to 24x 24. The primary classifier and the secondary classifier are trained offline using positive and negative sample sets.
In this embodiment, the first-stage classifier is an AdaBoost classifier based on a cascade model. The training process calculates Haar _ like characteristics of positive and negative samples to train the AdaBoost classifier. Namely, an integral graph of the sample feature is rapidly calculated by using edge features and linear features in the Haar _ like feature, and then the Haar _ like feature is calculated in the form of the integral graph. Since the feature for Haar _ like is a conventional feature of an image, the present invention does not need to describe the calculation and acquisition process in detail. The invention briefly describes an AdaBoost classifier based on a cascade model, wherein the AdaBoost is used as a boosting method, each training sample is endowed with an equal initial weight at the beginning, then a weak classifier is trained on a training set, the error rate of the weak classifier is calculated, the weight of each sample in the training set is readjusted according to the obtained error rate of the weak classifier, so that the classified error samples obtain higher weight and the classified correct samples reduce weight. After the weight of each sample is updated, the next round of training is started until the set iteration number is reached or the error rate of the classifier is zero, and then a strong classifier is combined according to the weights of the weak classifiers. The invention takes the classifier trained by the AdaBoost algorithm as a strong classifier in the cascade model. Each strong classifier has high judgment accuracy on negative samples, when a strong classifier at one stage judges that the negative samples exist, the classifier at the next stage is not called continuously, most of negative samples exist in one image under the condition of a traffic scene, and a large number of negative samples can be abandoned at the initial stage of the cascade classifier by the method, so that the detection speed of the classifier is increased. Meanwhile, only the previous classifier in the cascade classifier judges that the sample is positive, the sample is sent to the next classifier for judging again, and the false detection rate of the vehicle candidate area is reduced in the vehicle rough detection stage.
In summary, the method utilizes the AdaBoost classifier based on the cascade model to perform vehicle detection on the video frame image acquired in real time to acquire the candidate region of the target vehicle. For example, when the current frame image is subjected to vehicle detection, the scanning window sends the image in the window into a first-level classifier, so that whether the image in the window contains a vehicle or not is identified, and if the image contains the vehicle, the image is a target vehicle candidate area; otherwise, discarding.
In this embodiment, the secondary classifier is selected as a support vector machine, which calculates HOG features through positive and negative sample sets to obtain a HOG feature descriptor, and sends the calculated HOG feature vector as an input vector to a Support Vector Machine (SVM) classifier for training. Wherein, the size of the detection window is 56x56, the block size is 16x16, the cell size is 8x8, the block sliding step is 8x8, and the number of bins is set to 9. Similarly, since the HOG feature is a conventional feature of an image, the calculation and acquisition process thereof will not be described in detail in the present invention. The secondary classifier constructed by the invention is used for further identifying the classification result of the primary classifier in practical application.
In summary, the present invention uses the first-stage classifier to classify and identify the current frame image to obtain the target vehicle candidate region, and then uses the second-stage classifier to classify and identify the target vehicle candidate region to obtain the target vehicle region on the current frame image.
S3: based on the target vehicle region extracted in step S2, each kernel-dependent filtering tracker tracks the target vehicle based on the target vehicle region in the current frame image and other frame images in the video data to obtain tracking information of the corresponding target vehicle.
First, a kernel correlation filter tracker (KCF) used in the present invention is described, where the kernel correlation filter tracker (KCF) is a single target tracker, the tracking process is to initially define a certain target in a video stream, and then an algorithm is used to find a region closest to the initial target in the next frame of image, and the region is considered as a tracking region of the target in the current frame. The invention selects the KCF algorithm.
Secondly, the invention constructs a multi-target tracking management module for tracking the target vehicle. The multi-target tracking management module is designed aiming at the condition that generally more than one target vehicle exists in a traffic scene, and integrates the tracking of a plurality of targets and the timely updating of the targets. One core-related filtering tracker tracks one target vehicle, and all current core-related filtering trackers and corresponding target vehicles form a current tracking queue. Therefore, the same target vehicle is tracked in the video stream by the same kernel correlation filter, which utilizes the video stream images to determine the tracking area of the target vehicle in the video stream.
In order to make the target vehicle area detected by other frame images in the video stream be the corresponding area of the new target vehicle, the following operations are executed when the next frame image is used in the invention:
firstly, masking a tracking area where a target vehicle in a current tracking queue in a next frame of image is located; then, the masked next frame image is processed according to step S2 and step S3.
And acquiring a tracking area of the target vehicle in the next frame of image by using a nuclear correlation filtering tracker of the target vehicle in the current tracking queue, and setting all pixel values of the corresponding tracking area to be 255. Therefore, when the next frame of image is processed according to the steps S2 and S3, the identified target vehicles are all new target vehicles, namely are not in the current tracking queue, and a kernel correlation filter tracker can be assigned to track the new target vehicles. According to the invention, the new target area can be rapidly identified through the method, and the area after mask processing is rapidly judged as a negative sample in the primary classifier, namely, the area does not contain vehicles.
It should be understood that the present invention processes the other frame images in sequence according to steps S2 and S3, and at the same time, the nuclear correlation filtering tracker synchronously performs tracking to acquire the tracking areas of the corresponding target vehicles in the other frame images.
In order to enable the nuclear correlation filtering tracker to have a better tracking effect, the nuclear correlation filter is improved, each target vehicle is provided with an online monitor, and the online monitors can judge whether the target vehicles disappear in a visual field range or reach the boundary position of a retrograde detection area. For example, if the upper left corner coordinate or the lower right corner coordinate of the target detection frame of the target vehicle in the continuous frame images meets the set boundary threshold, it can be determined that the target vehicle has reached the boundary position of the reverse detection area, and the target vehicle is removed from the tracking queue.
In addition, the invention also trains a vehicle classifier off line for identifying the tracking effect of the nuclear correlation filter and judging whether the tracked target drifts or not. The vehicle classifier is the same as the first-stage classifier and the second-stage classifier and is used for identifying whether an image region contains a vehicle, the off-line training process also utilizes positive and negative samples for training, specifically, gradient information of each pixel point in the positive and negative samples is utilized, the gradient direction is divided into 9 bins, the number of the pixel points in each bin is counted, a statistic histogram of the gradient direction of the positive and negative samples is obtained, the statistic histogram is a feature vector of the positive and negative positive samples, and the feature vector is utilized for off-line training.
In the embodiment, when the tracking effect of a certain nuclear correlation filter tracker is identified, a statistical histogram of the gradient direction of a tracking area is calculated based on the tracking area detected by a nuclear correlation filter in the next frame, and then whether the tracking area has a vehicle is identified by a vehicle classifier, and if the tracking area has a vehicle, the tracking area is considered to be correct; if no vehicle is present, the tracking is considered to be in error.
In the invention, each kernel related filtering tracker adopts the method to carry out punishment scoring, when a vehicle classifier obtains the probability of a vehicle as p every time an error occurs, the score is 1-p, and if the accumulation reaches S, the score is removed from a tracking queue.
S4: whether the target vehicle is in a wrong-way driving detection area or not is discriminated and the driving state of the target vehicle is detected based on the tracking information of the target vehicle and the wrong-way driving direction mark.
In this embodiment, the tracking process of a target vehicle in consecutive frame images will obtain the target detection frame shown in fig. 3. And recording the coordinate value of the lower center position of the target detection frame, and further obtaining the motion track growth direction and the motion track growth length of the target vehicle. In this embodiment, it is preferable to record the coordinate value of the lower center position of the target detection frame every n1 frames, thereby reducing space waste and preventing data redundancy.
The invention identifies whether the target vehicle drives in the wrong direction or not based on the motion trail growth direction of the target vehicle and identifies the running state of the target vehicle based on the motion trail growth length of the target vehicle, wherein the identification rule is as follows:
a about retrograde identification:
judging whether the motion trail growth direction is the same as the retrograde motion direction, if so, performing retrograde motion corresponding to the target vehicle; if not, no retrograde motion occurs.
B: regarding the travel state discrimination:
firstly, acquiring the motion track growth length of the lower central point of a target detection frame based on the position of the target detection frame of a target vehicle in a continuous N1 frame image; then, judging whether the growth length of the motion trail is greater than a preset first threshold value L1;
if the speed is larger than the set speed and the reverse driving does not occur, the target vehicle normally runs;
otherwise, acquiring a new motion track growth length of a lower central point of the target detection frame based on the position of the target detection frame of the target vehicle in the continuous N2 frame images, and judging whether the new motion track growth length is smaller than a preset second threshold value L2;
wherein N2 is greater than N1, and L2 is less than L1;
if the current time interval is smaller than a preset second threshold value L2, judging whether other target vehicles exist in the same driving direction at the current same time interval;
if other target vehicles exist, acquiring the motion track growth lengths of the other target vehicles in the same driving direction in the same time period in the continuous N1 frames;
if the motion track growth lengths of other target vehicles in the continuous N1 frames are greater than the preset first threshold value L1, the target vehicles are in an abnormal state, such as a fault parking. If the motion trail growth lengths of other target vehicles in the continuous N1 frames are all smaller than a preset second threshold value L2, the target vehicles are in a blocked state; if the motion trail growth lengths of other targets in the continuous N1 frames are less than or equal to a preset first threshold L1 and greater than or equal to a second threshold L2, the target vehicle is in a pending state;
if no other target vehicle exists, the target vehicle is in an abnormal state;
and if the growth length of the new motion trail is not less than a preset second threshold value L2, the target vehicle is in a pending state. Wherein the pending state is a state requiring manual authentication.
In conclusion, the invention provides a means for automatically identifying the reverse running of the vehicle through the above mode, and simultaneously identifies the running state of the vehicle, thereby expanding the application range of the vehicle.
It should be emphasized that the examples described herein are illustrative and not restrictive, and thus the invention is not to be limited to the examples described herein, but rather to other embodiments that may be devised by those skilled in the art based on the teachings herein, and that various modifications, alterations, and substitutions are possible without departing from the spirit and scope of the present invention.
Claims (7)
1. An intelligent detection method for vehicle reverse running based on tracking trajectory analysis is characterized in that: the method comprises the following steps:
s1: acquiring a retrograde motion detection area and a retrograde motion direction mark on a shooting area image of a camera and acquiring video data of the camera in real time;
the retrograde motion detection area is a quadrilateral area containing a vehicle driving lane, and the retrograde motion direction is marked as a directional line segment;
s2: extracting a target vehicle region in a reverse detection region from a current frame of image of video data based on a first-stage classifier and a second-stage classifier of off-line training;
the method comprises the steps that a first-stage classifier is used for classifying and identifying a current frame image to obtain a target vehicle candidate region, and a second-stage classifier is used for classifying and identifying the target vehicle candidate region to obtain a target vehicle region on the current frame image;
the input data of the first-stage classifier and the second-stage classifier in the off-line training process are respectively a positive sample image and a negative sample image, wherein the positive sample image contains vehicles, and the negative sample image does not contain vehicles;
s3: based on the target vehicle region extracted in step S2, distributing relevant filtering trackers, each kernel relevant filtering tracker tracking the target vehicle based on the target vehicle region in the current frame image and other frame images in the video data to obtain tracking information of the corresponding target vehicle;
a core related filtering tracker tracks a target vehicle, and all current core related filtering trackers and corresponding target vehicles form a current tracking queue;
sequentially processing other frame images in the video data according to the step S2 and the step S3, wherein if an identified target vehicle is not in the current tracking queue in the processing process, the target vehicle is a new target vehicle, and a kernel-related filtering tracker is distributed for tracking the new target vehicle;
s4: identifying whether the target vehicle has retrograde motion in the retrograde motion detection area based on the tracking information of the target vehicle and the retrograde motion direction mark;
the tracking information comprises target detection frame positions of the target vehicle in the continuous frame images, and the target detection frames represent tracking areas of the target vehicle in each frame image; acquiring the growth direction of a motion trail at the same position on a target detection frame based on the position of the target detection frame of a target vehicle in a continuous frame image, judging whether the growth direction of the motion trail is the same as the retrograde direction, and if so, performing retrograde motion corresponding to the target vehicle; if not, no retrograde motion occurs;
when the tracking information of the corresponding target vehicle is obtained by using the nuclear correlation filtering tracker, punishment scoring is carried out on the tracking of the nuclear correlation filtering tracker;
acquiring a tracking area of a target vehicle in a next frame of image based on a kernel correlation filtering tracker, and identifying whether the tracking area contains the vehicle by using a vehicle classifier trained offline in advance; if the vehicle is not included, the vision-check related filtering tracker makes a tracking error, and punishment scoring is carried out on the tracking error;
when the probability p is greater than or equal to 0.5, the vehicle is included, and when the probability p is less than 0.5, the vehicle is not included; when the vehicle classifier obtains that the probability of the vehicle is p, the score is 1-p;
if the accumulated score of the kernel-related filtering tracker reaches a preset S score, deleting the kernel-related filtering tracker in a current tracking queue;
the input data of the vehicle classifier training process are positive sample images and negative sample images.
2. The method of claim 1, wherein: if the vehicle does not run backwards, the method also comprises the step of identifying the running state of the target vehicle based on the tracking information of the target vehicle, and the process is as follows:
firstly, acquiring the motion trail growth length of the same position on a target detection frame based on the position of the target detection frame of a target vehicle in continuous N1 frame images; then, judging whether the growth length of the motion trail is greater than a preset first threshold value L1;
if the current speed is greater than a preset first threshold value L1, the target vehicle runs normally;
otherwise, acquiring the growth length of a new motion track at the same position on the target detection frame based on the position of the target detection frame of the target vehicle in the continuous N2 frame images, and judging whether the growth length of the new motion track is smaller than a preset second threshold value L2;
wherein N2 is greater than N1, and L2 is less than L1;
if the current time interval is smaller than a preset second threshold value L2, judging whether other target vehicles exist in the same driving direction at the current same time interval;
if other target vehicles exist, acquiring the motion track growth lengths of the other target vehicles in the same driving direction in the same time period in the continuous N1 frames; if the motion track growth lengths of other target vehicles in the continuous N1 frames are greater than a preset first threshold value L1, the target vehicles are in an abnormal state; if the motion trail growth lengths of other target vehicles in the continuous N1 frames are all smaller than a preset second threshold value L2, the target vehicles are in a blocked state; if the motion trail growth lengths of other targets in the continuous N1 frames are less than or equal to a preset first threshold L1 and greater than or equal to a second threshold L2, the target vehicle is in a pending state;
if no other target vehicle exists, the target vehicle is in an abnormal state;
and if the growth length of the new motion trail is not less than a preset second threshold value L2, the target vehicle is in a pending state.
3. The method of claim 2, wherein: the value ranges of the preset first threshold value L1 and the preset second threshold value L2 corresponding to the target vehicle are respectively as follows: l1>2/3h, L2<1/3 h;
and h is the high pixel value of the target detection frame when the tracker captures the target vehicle for the first time.
4. The method of claim 2, wherein: in step S4, the motion trajectory growth direction and length of the lower center point position of the target detection frame are obtained based on the position of the target detection frame of the target vehicle in the continuous frame images;
wherein, every interval n1 frames records the coordinate value of the lower center point position of the target detection frame.
5. The method of claim 1, wherein: when processing the next frame image, the method also comprises the following steps:
firstly, masking a tracking area where a target vehicle in a current tracking queue in a next frame of image is located;
secondly, acquiring a tracking area of the target vehicle in the next frame of image by using a nuclear correlation filtering tracker of the target vehicle in the current tracking queue;
then, the next frame image after masking is processed according to steps S2 and S3, where the masking is to set the pixel values of the pixel points corresponding to the tracking areas to 255.
6. The method of claim 1, wherein: when the tracking information of the corresponding target vehicle is acquired by using the nuclear correlation filtering tracker, whether the target vehicle disappears in a visual field range or reaches the boundary position of a retrograde detection area is supervised;
and if the current tracking queue disappears in the visual field range or the target vehicle reaches the boundary position of the reverse driving detection area, deleting the kernel-related filtering tracker in the current tracking queue.
7. The method of claim 1, wherein: the first-stage classifier is an AdaBoost classifier based on a cascade model; the secondary classifier is a support vector machine;
training the training process of the primary classifier based on Haar _ like characteristics of positive samples and negative samples;
the training process of the secondary classifier is trained on the basis of feature vectors formed by HOG features of positive samples and negative samples.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910244210.8A CN109948582B (en) | 2019-03-28 | 2019-03-28 | Intelligent vehicle reverse running detection method based on tracking trajectory analysis |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910244210.8A CN109948582B (en) | 2019-03-28 | 2019-03-28 | Intelligent vehicle reverse running detection method based on tracking trajectory analysis |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109948582A CN109948582A (en) | 2019-06-28 |
| CN109948582B true CN109948582B (en) | 2021-03-02 |
Family
ID=67012282
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910244210.8A Active CN109948582B (en) | 2019-03-28 | 2019-03-28 | Intelligent vehicle reverse running detection method based on tracking trajectory analysis |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109948582B (en) |
Families Citing this family (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110472496B (en) * | 2019-07-08 | 2022-10-11 | 长安大学 | Traffic video intelligent analysis method based on target detection and tracking |
| CN110555423B (en) * | 2019-09-09 | 2021-12-21 | 南京东控智能交通研究院有限公司 | Multi-dimensional motion camera-based traffic parameter extraction method for aerial video |
| CN110473405B (en) * | 2019-09-25 | 2020-10-23 | 拉扎斯网络科技（上海）有限公司 | Driving state detection method and device, readable storage medium and electronic equipment |
| CN111126171A (en) * | 2019-12-04 | 2020-05-08 | 江西洪都航空工业集团有限责任公司 | Vehicle reverse running detection method and system |
| CN111046797A (en) * | 2019-12-12 | 2020-04-21 | 天地伟业技术有限公司 | Oil pipeline warning method based on personnel and vehicle behavior analysis |
| CN111354021A (en) * | 2020-02-14 | 2020-06-30 | 广东工业大学 | Target tracking method based on target identification and pixel marking |
| CN113327414B (en) * | 2020-02-28 | 2023-06-20 | 深圳市丰驰顺行信息技术有限公司 | Vehicle reverse running detection method, device, computer equipment and storage medium |
| CN111259868B (en) * | 2020-03-10 | 2023-12-12 | 以萨技术股份有限公司 | Reverse vehicle detection method, system and medium based on convolutional neural network |
| CN111833598B (en) * | 2020-05-14 | 2022-07-05 | 山东科技大学 | Automatic traffic incident monitoring method and system for unmanned aerial vehicle on highway |
| CN111611901B (en) * | 2020-05-15 | 2023-10-03 | 北京百度网讯科技有限公司 | Vehicle reverse running detection method, device, equipment and storage medium |
| CN112820115B (en) * | 2021-01-05 | 2021-11-23 | 四川铁投信息技术产业投资有限公司 | Method for identifying steering state of running vehicle |
| CN112686215B (en) * | 2021-01-26 | 2023-07-25 | 广东工业大学 | Track tracking monitoring and early warning system and method for carrier vehicle |
| CN112507993B (en) * | 2021-02-05 | 2021-06-15 | 上海闪马智能科技有限公司 | Abnormal driving-away detection method and system, electronic equipment and storage medium |
| CN113570877B (en) * | 2021-06-22 | 2022-09-23 | 淮阴工学院 | Non-motor vehicle retrograde detection device and detection method |
| CN117523521B (en) * | 2024-01-04 | 2024-04-02 | 山东科技大学 | Vehicle detection method based on Haar features and improved HOG features |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102768804A (en) * | 2012-07-30 | 2012-11-07 | 江苏物联网研究发展中心 | Video-based traffic information acquisition method |
| CN105447511A (en) * | 2015-11-15 | 2016-03-30 | 安徽清新互联信息科技有限公司 | SVM object detection method based on Adaboost Haar-Like features |
| WO2016080913A1 (en) * | 2014-11-18 | 2016-05-26 | Agency For Science, Technology And Research | Method and device for traffic sign recognition |
| CN106570439A (en) * | 2015-10-09 | 2017-04-19 | 腾讯科技（深圳）有限公司 | Vehicle detection method and apparatus |
| CN109190523A (en) * | 2018-08-17 | 2019-01-11 | 武汉大学 | A kind of automobile detecting following method for early warning of view-based access control model |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104809887B (en) * | 2015-04-23 | 2018-03-16 | 杭州中威电子股份有限公司 | A kind of retrograde detection method of vehicle on expressway and autoalarm |
| US9613273B2 (en) * | 2015-05-19 | 2017-04-04 | Toyota Motor Engineering & Manufacturing North America, Inc. | Apparatus and method for object tracking |
| CN105354529A (en) * | 2015-08-04 | 2016-02-24 | 北京时代云英科技有限公司 | Vehicle converse running detection method and apparatus |
| WO2018049514A1 (en) * | 2016-09-13 | 2018-03-22 | Genetec Inc. | Auto-calibration of tracking systems |
| CN106355884B (en) * | 2016-11-18 | 2019-02-12 | 成都通甲优博科技有限责任公司 | A kind of vehicle on highway guidance system and method based on vehicle classification |
| KR102434397B1 (en) * | 2017-07-10 | 2022-08-22 | 한국전자통신연구원 | Real time multi-object tracking device and method by using global motion |
-
2019
- 2019-03-28 CN CN201910244210.8A patent/CN109948582B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102768804A (en) * | 2012-07-30 | 2012-11-07 | 江苏物联网研究发展中心 | Video-based traffic information acquisition method |
| WO2016080913A1 (en) * | 2014-11-18 | 2016-05-26 | Agency For Science, Technology And Research | Method and device for traffic sign recognition |
| CN106570439A (en) * | 2015-10-09 | 2017-04-19 | 腾讯科技（深圳）有限公司 | Vehicle detection method and apparatus |
| CN105447511A (en) * | 2015-11-15 | 2016-03-30 | 安徽清新互联信息科技有限公司 | SVM object detection method based on Adaboost Haar-Like features |
| CN109190523A (en) * | 2018-08-17 | 2019-01-11 | 武汉大学 | A kind of automobile detecting following method for early warning of view-based access control model |
Non-Patent Citations (2)
| Title |
|---|
| Robust online multi-object tracking based on KCF trackers and reassignment;Huiling Wu等;《2016 IEEE Global Conference on Signal and Information Processing》;20161209;第124-128页 * |
| 一种基于多特征融合级联分类器的车辆检测算法;周行等;《现代计算机(专业版)》;20180620;第38-43页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109948582A (en) | 2019-06-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109948582B (en) | Intelligent vehicle reverse running detection method based on tracking trajectory analysis | |
| CN108053427B (en) | Improved multi-target tracking method, system and device based on KCF and Kalman | |
| CN108009473B (en) | Video structuralization processing method, system and storage device based on target behavior attribute | |
| CN108062349B (en) | Video monitoring method and system based on video structured data and deep learning | |
| CN108052859B (en) | Abnormal behavior detection method, system and device based on clustering optical flow characteristics | |
| Zhang et al. | Application of deep learning and unmanned aerial vehicle technology in traffic flow monitoring | |
| Bach et al. | Deep convolutional traffic light recognition for automated driving | |
| CN104134078B (en) | Automatic selection method for classifiers in people flow counting system | |
| CN109829382B (en) | Abnormal target early warning tracking system and method based on intelligent behavior characteristic analysis | |
| CN104915655A (en) | Multi-path monitor video management method and device | |
| CN104751136A (en) | Face recognition based multi-camera video event retrospective trace method | |
| Li et al. | Robust people counting in video surveillance: Dataset and system | |
| WO2011097795A1 (en) | Method and system for population flow statistics | |
| CN102542289A (en) | Pedestrian volume statistical method based on plurality of Gaussian counting models | |
| Chang et al. | Video analytics in smart transportation for the AIC'18 challenge | |
| CN103116987A (en) | Traffic flow statistic and violation detection method based on surveillance video processing | |
| CN102663452A (en) | Suspicious act detecting method based on video analysis | |
| Pan et al. | Traffic surveillance system for vehicle flow detection | |
| CN102254428B (en) | Traffic jam detection method based on video processing | |
| CN103246896A (en) | Robust real-time vehicle detection and tracking method | |
| KR101472674B1 (en) | Method and apparatus for video surveillance based on detecting abnormal behavior using extraction of trajectories from crowd in images | |
| CN111898485A (en) | Parking space vehicle detection processing method and device | |
| CN114898326A (en) | Method, system and equipment for detecting reverse running of one-way vehicle based on deep learning | |
| CN116153086B (en) | Multi-path traffic accident and congestion detection method and system based on deep learning | |
| CN108520528B (en) | Mobile vehicle tracking method based on improved difference threshold and displacement matching model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |