CN112749210A

CN112749210A - Vehicle collision recognition method and system based on deep learning

Info

Publication number: CN112749210A
Application number: CN202110061524.1A
Authority: CN
Inventors: 朱瑞; 何道峰
Original assignee: Ubiai Information Technology Beijing Co ltd
Current assignee: Ubiai Information Technology Beijing Co ltd
Priority date: 2021-01-18
Filing date: 2021-01-18
Publication date: 2021-05-04
Anticipated expiration: 2041-01-18
Also published as: CN112749210B

Abstract

The invention relates to a vehicle collision recognition method and a system based on deep learning, wherein the vehicle collision recognition method comprises the following steps: step 1: acquiring real-time running data of a vehicle, and acquiring historical running data and historical danger data of the vehicle; step 2: preprocessing vehicle operation data; and step 3: performing low-frequency to high-frequency data interpolation on the vehicle operation data; and 4, step 4: ETL engineering characteristic transformation; and 5: performing model training and iteration through a deep learning method and a mathematical statistics correlation principle to obtain a correlation model; step 6: and obtaining a vehicle collision recognition result according to the real-time vehicle operation data and by combining the relevant models. The method adopts massive Internet of vehicles data, models 5 angles from abnormal running track, abnormal deceleration, brake distance prediction, abnormal parking and abnormal turning, and identifies collision accidents, thereby avoiding judgment troubles caused by complex geographic factors.

Description

Vehicle collision recognition method and system based on deep learning

Technical Field

The invention relates to the technical field of vehicle collision detection, in particular to a vehicle collision identification method and system based on deep learning.

Background

With the rapid increase of the number of automobiles in China, road traffic accidents are frequently happened, so that serious casualties and property losses are caused, the road traffic safety situation is severe, and the attention of the whole society is aroused. A large number of surveys show that after a road traffic accident occurs, particularly after a collision accident occurs, the number of casualties caused by untimely rescue is far larger than that caused by the accident, and therefore, the vehicle collision accident can be timely known, and the rescue requirement can be obtained. Due to the limitation of data collection conditions, in the prior art, the method for detecting the collision mainly uses a threshold value of a gravity acceleration sensor as a judgment condition for detecting the collision. Many current detection methods read the values of the three axes of the acceleration sensor X, Y, Z by polling, calculate the difference between the maximum value and the minimum value of each axis, and then calculate the vector of the three axes. By comparing this vector value with a predetermined value, a collision is considered to have occurred if the vector value is greater than the predetermined value. However, when the vehicle is driven on a complex road condition, the detection technology can cause the situation of false alarm or false alarm. For example, suppose the road conditions have sewer well lid or deceleration strip, the vehicle can take place great vibrations when passing these places, and some car owners have great brake dynamics when passing these places for the acceleration sensor who installs on the car produces great difference, has just produced the wrong report very easily. Meanwhile, because the parameters of the bodies of vehicles of different models are different and the collision preset values of different products are also different, the collision detection technology is very difficult to realize in use and maintenance aiming at various vehicle models and various application scenes, the accuracy of an output result is not enough, a large amount of misjudgments or missed judgments can be caused, and correct judgments can not be made in time when an accident occurs. Some collision detection methods, such as an OBD interface-based vehicle-mounted equipment judgment method, have the disadvantages that the method is not compatible with all vehicle types, the equipment cost is high, the detection algorithm is simple, and the number of collision scenes is small; the method for judging the vehicle-mounted equipment based on the video images has the defects of high equipment manufacturing cost, easy damage of the equipment in the collision process, low cost performance of detection and calculation, less coverage on collision scenes and low detection precision.

Disclosure of Invention

In view of the above, the present invention provides a method and a system for recognizing a vehicle collision based on deep learning, which overcome the disadvantages of the prior art.

In order to achieve the purpose, the invention adopts the following technical scheme: a vehicle collision recognition method based on deep learning comprises the following steps:

step 1: acquiring real-time running data of a vehicle, and acquiring historical running data and historical danger data of the vehicle;

step 2: preprocessing vehicle operation data;

and step 3: performing low-frequency to high-frequency data interpolation processing on the vehicle operation data;

and 4, step 4: ETL engineering characteristic conversion treatment;

and 5: performing model training and iteration through a deep learning method and a mathematical statistics correlation principle to obtain a correlation model;

step 6: and obtaining a vehicle collision recognition result according to the real-time vehicle operation data and by combining the relevant models.

Optionally, the acquiring historical vehicle operation data and historical vehicle risk data includes:

obtaining historical vehicle operating data, the data comprising: a timestamp, a GPS speed, a GPS direction, a GPS altitude, a device power status, a device location status, a vehicle movement status, a GPS longitude, and a GPS latitude;

acquiring historical insurance information of the vehicle through a third party, wherein the insurance information comprises: the time of the accident, the place of the accident and the collision position of the vehicle;

matching the insurance information with the historical vehicle running data through insurance time, intercepting time sequence data with different time length Ts at the insurance moment by a data sliding method, marking the data as insurance, and simultaneously intercepting a certain amount of non-insurance time sequence data and marking the data as non-insurance.

Optionally, the preprocessing the vehicle operation data includes:

carrying out time sequence chaotic processing on vehicle operation data;

processing the abnormal value of GPS positioning drift of the data after the time sequence disorder processing;

and (3) performing signal loss data processing on the data subjected to the GPS positioning drift abnormal value processing, and deleting the data which has no positioning signal and has the speed of 0.

Optionally, the performing time-series chaotic processing on the vehicle operation data includes:

for historical data used for model training, performing mode calculation on min records of the Ts time sequence data, comparing each record with the current mode, calculating a timestamp difference, and if the difference is greater than preset time, considering that the record is a complement operation on previous data, and deleting the record;

for the time sequence data calculated in real time on line of the model, calculating the time difference between the time stamp of each piece of data and the current system time, if the difference is greater than the preset time, considering the piece of data as the operation of system supplementary recording data, and deleting the piece of data;

and sequencing the processed historical data or the real-time analyzed data according to the time stamp.

Optionally, the processing of the GPS positioning drift outlier on the data after the time-series chaos processing includes:

sequentially calculating the time interval between the t and t +1 records and the earth linear distance;

obtaining a suspicious straight line distance;

and positioning the drift point according to the suspicious linear distance, the time interval and the earth linear distance.

Optionally, the performing low-frequency to high-frequency data interpolation processing on the vehicle operation data includes:

carrying out low-frequency to high-frequency data interpolation processing on vehicle operation data by using a differential integral method and a Bezier curve method;

specifically, a differential integration method is used for carrying out speed term interpolation processing on vehicle running data;

and carrying out GPS longitude and latitude item interpolation on the vehicle running data subjected to the speed item interpolation processing by using a Bezier curve method.

Optionally, the model training and iteration performed by the deep learning method and the mathematical statistics correlation principle includes:

performing training set and test set division on the processed historical data;

and performing LSTM modeling on the training set data by using LSTM model parameters, and evaluating the test set data prediction result by using LSTM model optimization and evaluation parameters so as to perform optimization iteration on the model parameters.

Optionally, the correlation model includes: the system comprises a running track abnormity identification model, an emergency brake abnormity identification model, a turning abnormity identification model, a parking abnormity identification model and a brake distance abnormity identification model;

the obtaining of the recognition result according to the real-time vehicle operation data and the combination of the relevant models comprises the following steps:

dividing the real-time vehicle running data according to a time interval of Ts;

processing the divided Ts duration data in real time according to the steps 2, 3 and 4;

inputting the processed data into a running track abnormity identification model, a sudden braking abnormity identification model, a turning abnormity identification model, a parking abnormity identification model and a braking distance abnormity identification model to respectively obtain driving abnormity scores of the models;

and determining a recognition result according to the driving abnormity scores of the models.

Optionally, the determining the recognition result according to the driving abnormality score of each model includes:

solving the sum of the products of the abnormal driving scores of the models and the corresponding weights of the models, and taking the sum as a prediction score;

judging whether the prediction score is greater than an accident threshold value;

if the predicted score is larger than the accident threshold value, judging that the vehicle has a collision accident; otherwise, judging that the vehicle has no collision accident.

The invention also provides a vehicle collision recognition system based on deep learning, which comprises:

the data acquisition module is used for acquiring real-time running data of the vehicle and acquiring historical running data and historical danger data of the vehicle;

the preprocessing module is used for preprocessing vehicle operation data;

the data interpolation processing module is used for performing low-frequency to high-frequency data interpolation processing on the vehicle operation data;

the conversion module is used for ETL engineering characteristic conversion treatment;

the modeling module is used for carrying out model training and iteration through a deep learning method and a mathematical statistics correlation principle to obtain a correlation model;

and the recognition module is used for obtaining a vehicle collision recognition result according to the processed vehicle real-time operation data and by combining the relevant models.

By adopting the technical scheme, the vehicle collision recognition method based on deep learning comprises the following steps: step 1: acquiring real-time running data of a vehicle, and acquiring historical running data and historical danger data of the vehicle; step 2: preprocessing vehicle operation data; and step 3: performing low-frequency to high-frequency data interpolation processing on the vehicle operation data; and 4, step 4: ETL engineering characteristic conversion treatment; and 5: performing model training and iteration through a deep learning method and a mathematical statistics correlation principle to obtain a correlation model; step 6: and obtaining a vehicle collision recognition result according to the real-time vehicle operation data and by combining the relevant models. According to the identification method, mass Internet of vehicles data are adopted, so that the accuracy of the judgment result is improved; factors such as driving behavior factors, gender, weather conditions and the like of the car owner are added in the model establishing process, so that the judgment result is more accurate; the model identifies the collision accident from 5 angles of abnormal running track, abnormal deceleration, brake distance prediction, abnormal parking and abnormal turning, and the judgment angle is more comprehensive; the model judges whether a vehicle collision accident occurs or not from the result presentation of time sequence data and the change rule of the data, and replaces the mode that the traditional algorithm only uses one data record at a certain moment to judge, so that the judgment result is more reliable; the identification method can avoid judgment troubles caused by complex geographic factors when the traditional equipment is used for collision induction identification.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.

FIG. 1 is a schematic flow chart provided by a first embodiment of a deep learning-based vehicle collision recognition method of the present invention;

FIG. 2 is a schematic flow chart of a second embodiment of the deep learning-based vehicle collision recognition method according to the present invention;

FIG. 3 is a schematic view of a vehicle traveling at a constant speed;

FIG. 4 is a schematic view of a vehicle traveling at a non-uniform speed;

FIG. 5 is a schematic diagram of a first order Bezier curve interpolation GPS recording point;

FIG. 6 is a schematic diagram of a third order Bezier curve interpolation GPS recording point;

FIG. 7 is a schematic structural diagram provided by an embodiment of the deep learning-based vehicle collision recognition system of the present invention.

In the figure: 1. a data acquisition module; 2. a preprocessing module; 3. a data interpolation processing module; 4. a conversion module; 5. a modeling module; 6. and identifying the module.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be described in detail below. It is to be understood that the described embodiments are merely exemplary of the invention, and not restrictive of the full scope of the invention. All other embodiments, which can be derived by a person skilled in the art from the examples given herein without any inventive step, are within the scope of the present invention.

Fig. 1 is a schematic flow chart provided by a first embodiment of a deep learning-based vehicle collision recognition method according to the present invention.

As shown in fig. 1, a vehicle collision recognition method based on deep learning according to the embodiment includes:

step 2: preprocessing vehicle operation data;

and 4, step 4: ETL engineering characteristic conversion treatment;

In practical use, as shown in fig. 2, the vehicle collision recognition method described in this embodiment includes the following specific steps:

step 1: the method comprises the steps that real-time running data of a vehicle are collected through front-loading equipment and rear-loading equipment, and historical insurance data of the vehicle are obtained through a 4s dealer;

firstly, historical real-time running data of a vehicle is acquired, and the acquisition frequency is 1/5 seconds, wherein the acquisition frequency comprises the following steps: a timestamp, a GPS speed, a GPS direction, a GPS altitude, a device power status, a device location status, a vehicle movement status, a GPS longitude, and a GPS latitude.

Secondly, obtain historical emergence information through third party such as insurance company, dealer, car owner, including: time to accident, location to accident, and vehicle impact location.

And finally, matching the insurance information with the historical real-time running data of the vehicle through insurance time, intercepting different time sequence data with the time length Ts (which can be set to 120s) at the insurance moment by a data sliding method, marking the data as insurance, and meanwhile intercepting a certain amount of non-insurance time sequence data and marking the data as non-insurance.

Step 2: pre-processing vehicle operation data, comprising: the method comprises the following specific processes of time sequence disorder processing, GPS positioning drift abnormal value processing and signal loss data processing:

1) first, the time-series disorder processing: due to the influence of the defects of the data acquisition equipment, individual records cannot be acquired and transmitted to the interior of the system in real time, and are returned and stacked after the existing data after a certain time, so that the collision accident is misjudged by the model, and the 120s time sequence data needs to be screened and identified, wherein the method comprises the following steps:

for historical data used for model training, performing mode calculation on min records of 120s time sequence data, comparing each record with a current mode, calculating a timestamp difference, and if the difference is greater than preset time (such as 250s), considering the record as a device to perform complement operation on previous data, and processing the previous data by adopting a deletion method;

for the time sequence data calculated in real time on line of the model, calculating the time difference between the time stamp of each piece of data and the current system time, if the difference is greater than the preset time (such as 250s), considering the operation of supplementing the data by the system, and further deleting the piece of data; and finally, sequencing the processed historical data and the real-time analyzed data according to the time stamp.

2) Secondly, processing abnormal values of GPS positioning drift: due to the self-reason of the equipment, the problem of equipment positioning drift can be caused, so that the distance between two points in front of and behind the GPS is abnormally amplified, and the model result is misjudged, and therefore the equipment positioning drift judgment method comprises the following steps:

(1) sequentially calculating the time interval delta and the distance of the earth straight line between the t and t +1 records;

(2) obtaining the suspicious straight line distance (distance is more than 5000 m);

(3) sequentially judging the suspicious distance in the step (2), if the suspicious distance only has one item and the distance is formed by 120s time sequence data t₁₂₀And t₁₁₉If the calculated speed is larger than 40m/s, t is judged₁₂₀Recording as a positioning drift point, and deleting the record; if the suspicious distance is determined by t₁And t₂If the distance is calculated, the distance between the time interval delta corresponding to the travel and the next distance interval distance is judged_t23If Δ < 120s and distance_t23If < 1000m, t is determined₁For the drift point, if Δ < 120s and the next distance interval > 4000m, t is considered to be₂For the drift point, if the corresponding time interval Δ of the stroke is larger than 120s, and the distance is next_t23If the time interval is more than 4000m and the corresponding time interval of the next distance is less than 120s, the t is considered to be₂Is a drift point; if the suspicious distance is neither t₁And t₂Calculated as not t₁₂₀And t₁₁₉And calculating, if the time interval delta corresponding to the travel is less than 120s and the next distance interval is more than 4000m, considering that the second record in the distance is calculated as a drift point, and if the time interval delta corresponding to the travel is more than 120s and the next distance interval is more than 4000m and the time interval corresponding to the next distance is less than 120s, considering that the second record in the distance is calculated as a drift point.

3) And finally, processing the signal loss data, and deleting the data which has no positioning signal and has the speed of 0.

And step 3: and performing low-frequency to high-frequency data interpolation processing on the vehicle data by using a differential integration and Bezier curve method.

Because the data acquisition frequency of the equipment is sub/5 s, and the collision accident happens in a certain moment and is far less than 5s, the records need to be interpolated, the data frequency is improved by adopting the following method:

1) speed term interpolation: first, assuming that the vehicle is running at a constant speed between the records t and t +1 (as shown in fig. 3), the running distance in the time interval is calculated by using the integral principle, and the formula is as follows:

secondly, the earth straight-line distance k between the two points is calculated by using the corresponding GPS point, if s > k, the vehicle is not driven at a constant speed in the driving process, but is decelerated firstly and then accelerated (as shown in FIG. 4), so a record t3 is interpolated in the interval between t1 and t 2.

Two interpolation principles are specified:

the first one is: FIG. 4 the sum of the two trapezoidal areas equals s;

a second bar: the abnormality of the original data is not increased, and the two pieces of data are smooth as far as possible, so that the sum of the accelerations of the two strokes is required to be | a1| + | a2| as small as possible.

The calculation formula is as follows:

t＝t₁+t₂

since y is to be made ═ a₁|+|a₂L is minimized, so

the interpolation data t3 and v3 between t1 and t2 are as shown in the above results, and the other cases are analogized in turn.

2) Interpolation of longitude and latitude items of the GPS:

according to the scheme, the Bezier curve is utilized to approximately interpolate the GPS record between two GPS positioning points, and the method is specific:

if the GPS recording points are on the same straight line in 120s time sequence data, interpolating the GPS recording points by adopting a first-order Bezier curve, wherein the calculation formula is as follows:

B(t)＝(1-t)P₀+tP₁，t∈[0,1]

t is t1 in the velocity interpolation, as shown in FIG. 5.

If the GPS recording points are not on the same straight line in 120s time series data, a third-order Bezier curve is adopted to interpolate the GPS recording points, firstly, 4 GPS recording points G1(x1, y1), G2(x2, y2), G3(x3, y3) and G4(x4, y4) are assumed to exist,

Midpoint1_x＝(x1+x2)/2，Midpoint1_y＝(y1+y2)/2

Midpoint2_x＝(x2+x3)/2，Midpoint1_y＝(y2+y3)/2

Midpoint3_x＝(x3+x4)/2，Midpoint1_y＝(y3+y4)/2

K1＝d1/(d1+d2)

K2＝d2/(d2+d3)

split1＝midpoint1+K1*(midpoint2-midpoint1)

split2＝midpoint2+K2*(midpoint3-midpoint2)

translating the segment of midpoint1 and midpoint2 until split1 and G2 coincide, marking the coincident point as V1, and marking the translated points of midpoint1 and midpoint2 as V2 and V3, then

p0＝V2+K1*(V1-V2)

p1＝V3+K1*(V1-V3)

Similarly, p2, p3 can be calculated.

Note that G1, p0 is a control point between G1 and G2, p1 and p2 are control points between G2 and G3, and p2 and G4 are control points between G3 and G4.

The interpolated GPS record between two points can be calculated according to the following formula for third-order bezier curve smoothing, for example, the interpolation between two points AB is as follows:

M(t)＝(1-t)³A+3tB(1-t)²+3Ct²(1-t)+Dt³，t∈[0,1]

wherein, A and D are respectively a starting point and an end point, B and C are control points, and M (t) is an interpolated GPS recording point in A and D. t is t1 in the speed interpolation, as shown in FIG. 6.

And 4, step 4: ETL engineering feature transformation

And (3) performing characteristic transformation on the high-frequency data interpolated in the step (3), wherein the characteristic transformation is shown in table 1 and is specifically divided into the following characteristics:

TABLE 1

And 5: model training and iteration are carried out through a deep learning method and a mathematical statistics correlation principle, wherein the model comprises the following steps: the system comprises a running track abnormity identification model, an emergency brake abnormity identification model, a turning abnormity identification model, a parking abnormity identification model and a brake distance abnormity identification model.

1) The training and iteration process of the sudden braking abnormity model comprises the following steps:

firstly, taking real-time operation data of 1000 vehicles for 15 days, respectively calculating the absolute value of the negative acceleration of the vehicle at the current speed, and establishing a two-dimensional relation model of the speed-the absolute value of the negative acceleration;

secondly, segmenting the speed according to rounding, for example, v is 1m/s-2m/s, and finally calculating a 95% quantile (which can be set according to actual needs) of the negative acceleration absolute value corresponding to each speed interval as a threshold of deceleration abnormity;

finally, a relatively abnormal gear shifting behavior score p of the single deceleration behavior is calculated_i，p_iThe calculation method of (2) is as follows:

when a is_iWhen the value is greater than threshold, the score is recorded as 100; when a is_iWhen the value is less than or equal to the threshold, the score is recorded as p_i*100。

It should be noted that, 15-day real-time operation data of 1000 vehicles is taken as a sample, and the number of vehicles and the number of days of the sample can be set according to actual needs, which is only an example and does not limit the protection scope.

2) The training and iteration principle of the turning abnormality recognition model is the same as that of the emergency brake abnormality model.

3) The training and iteration process of the parking abnormity identification model comprises the following steps:

the parking abnormity is mainly recognized for the abnormal parking on the expressway, if the 120s time series data includes records with the speed greater than 110KM/h and records with the speed less than 3.6KM/h (considered parking records), and meanwhile, records of high-speed running exist in 10s before the vehicle is parked and the vehicle is not started in 120s time series after the vehicle is parked, the behavior is judged as the parking abnormity, and the abnormity score is recorded as 100.

4) The training and iteration process of the braking distance anomaly model comprises the following steps:

at the same speed, the minimum value of the braking distance of the vehicle is calculated according to the formula

d＝0.75*v+0.0256*v²；

Calculating the difference s between the actual deceleration distance and the minimum braking distance, and then calculating the abnormal braking distance score by the formula

5) The training and iteration process of the operation track abnormity identification model comprises the following steps:

and (4) carrying out deep learning LSTM model learning on the model data subjected to ETL engineering characteristic conversion in the step (4), wherein the LSTM (Long short term memory) principle is the prior art and is not described herein again.

Specifically, the steps of training the operation track abnormity identification model and predicting the model are as follows:

firstly, the historical data processed in the step 4 is divided into a training set and a test set.

Secondly, LSTM model building is carried out on the training set data by applying LSTM model parameters, and the test set data prediction result is evaluated by applying LSTM model optimization and evaluation parameters, so that optimization iteration is carried out on the model parameters.

And a sixth step: and (3) putting the model into use, collecting vehicle operation data in real time, processing the vehicle operation data collected in real time by using the steps 1, 2, 3 and 4, and inputting the processed vehicle operation data into the model in the step 5.

Specifically, the models are put into use, real-time data of the vehicle are processed in real time according to steps 1, 2, 3 and 4 at time intervals of 120s, then the processed data are input into 5 models in step 5, abnormal driving scores of the models are obtained respectively, weighting and summarizing are carried out according to certain weight to obtain predicted scores, and finally whether the vehicle has a collision accident or not is judged according to the size relation between the predicted scores and an accident threshold value. If the predicted score is larger than the accident threshold value, judging that the vehicle has a collision accident; otherwise, judging that the vehicle has no collision accident.

According to the identification method, mass Internet of vehicles data are adopted, so that the accuracy of the judgment result is improved; factors such as driving behavior factors, gender, weather conditions and the like of the car owner are added in the model establishing process, so that the judgment result is more accurate; the model identifies the collision accident from 5 angles of abnormal running track, abnormal deceleration, brake distance prediction, abnormal parking and abnormal turning, and the judgment angle is more comprehensive; the model judges whether a vehicle collision accident occurs or not from the result presentation of time sequence data and the change rule of the data, and replaces the mode that the traditional algorithm only uses one data record at a certain moment to judge, so that the judgment result is more reliable; the identification method can avoid judgment troubles caused by complex geographic factors when the traditional equipment is used for collision induction identification.

As shown in fig. 7, a deep learning based vehicle collision recognition system according to the present embodiment includes:

the data acquisition module 1 is used for acquiring real-time running data of the vehicle and acquiring historical running data and historical danger data of the vehicle;

the preprocessing module 2 is used for preprocessing vehicle operation data;

the data interpolation processing module 3 is used for performing low-frequency to high-frequency data interpolation processing on the vehicle operation data;

the conversion module 4 is used for ETL engineering characteristic conversion treatment;

the modeling module 5 is used for carrying out model training and iteration through a deep learning method and a mathematical statistics correlation principle to obtain a correlation model;

and the recognition module 6 is used for obtaining a vehicle collision recognition result according to the processed vehicle real-time operation data and by combining the relevant models.

The working principle of the vehicle collision recognition system based on deep learning in this embodiment is the same as that of the vehicle collision recognition method based on deep learning in any one of the above embodiments, and is not described herein again.

The vehicle collision recognition system based on deep learning provided by the invention adopts massive vehicle networking data, models 5 angles from abnormal running track, abnormal deceleration, braking distance prediction, abnormal parking and abnormal turning and recognizes collision accidents, avoids judgment troubles caused by complex geographic factors, and has the advantages of wide coverage on collision scenes and high detection precision.

It is understood that the same or similar parts in the above embodiments may be mutually referred to, and the same or similar parts in other embodiments may be referred to for the content which is not described in detail in some embodiments.

It should be noted that the terms "first," "second," and the like in the description of the present invention are used for descriptive purposes only and are not to be construed as indicating or implying relative importance. Further, in the description of the present invention, the meaning of "a plurality" means at least two unless otherwise specified.

Any process or method descriptions in flow charts or otherwise described herein may be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps of the process, and alternate implementations are included within the scope of the preferred embodiment of the present invention in which functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order, depending on the functionality involved, as would be understood by those reasonably skilled in the art of the present invention.

It should be understood that portions of the present invention may be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the various steps or methods may be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or combination of the following techniques, which are known in the art, may be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application specific integrated circuit having an appropriate combinational logic gate circuit, a Programmable Gate Array (PGA), a Field Programmable Gate Array (FPGA), or the like.

It will be understood by those skilled in the art that all or part of the steps carried by the method for implementing the above embodiments may be implemented by hardware related to instructions of a program, which may be stored in a computer readable storage medium, and when the program is executed, the program includes one or a combination of the steps of the method embodiments.

In addition, functional units in the embodiments of the present invention may be integrated into one processing module, or each unit may exist alone physically, or two or more units are integrated into one module. The integrated module can be realized in a hardware mode, and can also be realized in a software functional module mode. The integrated module, if implemented in the form of a software functional module and sold or used as a stand-alone product, may also be stored in a computer readable storage medium.

The storage medium mentioned above may be a read-only memory, a magnetic or optical disk, etc.

In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.

Claims

1. A vehicle collision recognition method based on deep learning is characterized by comprising the following steps:

step 2: preprocessing vehicle operation data;

and 4, step 4: ETL engineering characteristic conversion treatment;

2. The vehicle collision recognition method according to claim 1, wherein the acquiring vehicle historical operation data and vehicle historical emergence data includes:

3. The vehicle collision recognition method according to claim 2, wherein the preprocessing of the vehicle operation data includes:

carrying out time sequence chaotic processing on vehicle operation data;

4. The vehicle collision recognition method according to claim 3, wherein the subjecting of the vehicle operation data to time series confusion processing includes:

5. The vehicle collision recognition method according to claim 3, wherein the performing of the GPS positioning drift outlier processing on the time-series disorder processed data includes:

obtaining a suspicious straight line distance;

6. The vehicle collision recognition method according to claim 1, wherein the low-frequency to high-frequency data interpolation processing of the vehicle operation data includes:

7. The vehicle collision recognition method according to claim 1, wherein the model training and iteration through the deep learning method and the mathematical statistics correlation principle comprises:

performing training set and test set division on the processed historical data;

8. The vehicle collision recognition method according to any one of claims 2 to 7,

the correlation model includes: the system comprises a running track abnormity identification model, an emergency brake abnormity identification model, a turning abnormity identification model, a parking abnormity identification model and a brake distance abnormity identification model;

dividing the real-time vehicle running data according to a time interval of Ts;

9. The vehicle collision recognition method according to claim 8, wherein the determining a recognition result from the driving abnormality scores of the models includes:

10. A deep learning based vehicle collision recognition system, comprising:

the preprocessing module is used for preprocessing vehicle operation data;