A vehicle collision detection method and system based on feature time-domain matching

Abstract

The present invention provides a vehicle collision detection method and system based on feature time-domain matching. The method includes: S1: Using an adaptive data sampling algorithm to obtain high-frequency data packets uploaded in real time; S2: Performing local anomaly detection to determine whether there are local anomaly points. When there are local anomaly points, performing multi-dimensional feature processing on the data to obtain a feature set; S3: Performing feature matching through the constructed historical feature database to calculate the similarity distance; S4: Obtaining the minimum similarity distance and inputting it into the collision evaluation model to output the collision detection result. The system includes at least one device end, a collision detection end, and a historical feature database. The present invention collects data through an adaptive data collection algorithm, reduces the data collection pressure, extracts and fuses features in different dimensions, solves the problem of single-dimensional feature extraction, and performs matching detection through the construction of a historical feature database, thereby improving the accuracy of vehicle collision detection.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle collision detection, and in particular to a vehicle collision detection method and system based on feature time domain matching. Background Art

[0002] At present, with the increasing number of vehicles in my country, the negative impact of traffic problems is also gradually increasing. Among them, the loss of life and property and traffic congestion caused by traffic accidents have become urgent problems to be solved. While improving the supervision of drivers' traffic regulations, it is particularly important to establish real-time monitoring of vehicle collisions to quickly carry out accident rescue and traffic sorting. It is also of great significance to the construction of smart cities, which can enhance the perception of the urban environment and the intelligent regulation of urban traffic.

[0003] When a vehicle collides, the vehicle body often vibrates more significantly or produces a large acceleration. Therefore, the vehicle driving data is generally collected by relying on sensors installed on the vehicle to achieve real-time monitoring of the vehicle collision.

[0004] Since the collision time of a car accident is generally within the range of tens of milliseconds, there are high requirements for the sampling frequency of the sensor. The current sensor-based collision detection methods either have a low sampling frequency, sampling one data point per second or several seconds; or completely rely on high-frequency sampling data, but this will increase traffic costs and instability in system performance.

[0005] However, in the existing technology of vehicle collision detection based on high-frequency time series data, there are the following problems:

[0006] 1. The feature extraction dimension is relatively single, and the representation of data is not comprehensive enough. The existing technology basically relies on the features observed in a small number of samples by statistical methods or time-frequency transformation analysis methods, and there are strong assumptions about certain data distribution or human knowledge a priori. When the data becomes more complex and the amount of data gradually increases, the understanding of the data by this type of feature becomes more and more one-sided, and the assumptions made tend to be more and more random, and the collision detection effect is also poor.

[0007] 2. The existing technology generally focuses on anomaly detection of local data, ignoring the degree of outlier of the data point in the global data. However, the driving data of vehicles is highly complex and the data volume is large. Relying only on local anomaly detection methods will lead to a high number of false alarms of collisions and low detection accuracy.

[0008] 3. The distribution of driving data collected by different vehicle sensor devices varies greatly. Since the same feature value has different abnormality levels in the global data of different devices, there are large differences. Using the same global feature database to detect all vehicle collisions may lead to missed accidents. Summary of the invention

[0009] To solve the above problems, the present invention provides a vehicle collision detection method and system based on feature time domain matching, which adaptively collects data. Under normal circumstances, low-frequency sampling point data is performed, and high-frequency data sampling is performed when sensor data is abnormal, large acceleration or vehicle deceleration and stopping occurs. The collected high-frequency data is preprocessed twice to obtain a feature set, and an abnormal data feature library is constructed through the collected historical data. After the data of the feature set is obtained, it is compared with the data in the feature library, and whether to perform collision analysis is selected according to the comparison result, thereby improving the credibility of collision detection.

[0010] The present invention provides a vehicle collision detection method based on feature time domain matching, comprising:

[0011] S1: Adopts adaptive data sampling algorithm to obtain high-frequency data packets uploaded in real time;

[0012] S2: Perform local anomaly detection according to the high-frequency data packet to determine whether there is a local anomaly point. If there is a local anomaly point, perform multi-dimensional feature processing on the data to obtain a feature set, otherwise end;

[0013] S3: Based on the feature set, feature matching is performed with data in a pre-built historical feature database, similarity distance is calculated, and the historical data with the highest similarity in the historical feature database is obtained;

[0014] S4: Inputting the minimum similarity distance obtained by calculating and matching with the historical feature database into the collision assessment model to output the collision detection result.

[0015] Furthermore, in step S2, the multi-dimensional feature processing includes a first preprocessing and a second preprocessing, wherein the first preprocessing adopts different feature extraction methods to obtain multi-dimensional features, and the second preprocessing performs fusion and dimensionality reduction operations on the multi-dimensional features output by the first preprocessing to construct an output feature set.

[0016] Furthermore, the feature extraction method of the first preprocessing includes deep learning autoencoding, statistical method, and linear transformation.

[0017] Furthermore, the second preprocessing uses an RCA principal component analysis algorithm to fuse the different types of features obtained by the first preprocessing.

[0018] Furthermore, the historical feature database is constructed as follows:

[0019] Perform local anomaly detection on the equipment belonging to each vehicle to obtain relatively abnormal high-frequency data packets within a preset time period;

[0020] The obtained high-frequency data packets are subjected to feature processing, and a feature set of each abnormal high-frequency data packet is extracted and stored to obtain a historical feature database.

[0021] Furthermore, the adaptive data sampling algorithm is as follows:

[0022] Real-time detection of whether the sampling device or the vehicle collision detection device outputs a collision signal;

[0023] If no collision signal is output, low-frequency sampling is performed; if a collision signal is output, high-frequency sampling is performed.

[0024] In step S4, the collision detection result is specifically calculated as follows:

[0025] S401: Obtain the rank of the minimum similarity distance from large to small in the historical anomaly database and the quantile rank of the rank q , and the average speed avg_speed of the vehicle during the preset time period;

[0026] S402: The ranking rank and the corresponding quantile rank q , calculate the evaluation parameter size and calculate the collision detection score. The calculation formula is as follows:

[0027]

[0028]

[0029] Among them, a is the set proportion parameter, and its value is 10.

[0030] Furthermore, in step S2, after obtaining the real-time high-frequency data packet, the sensor data of each high-frequency sampling point in the trajectory segment of the predetermined duration is analyzed, and the analysis results of specific dimensions are vectorized respectively.

[0031] The present invention also provides a vehicle collision detection system based on feature time domain matching, comprising at least one device end, a collision detection end and a historical feature database, wherein the device end is data-connected to the collision detection end, and uploads data packets collected by the device end in real time, and the device end is provided with at least one collision detection device, and performs high-frequency sampling or low-frequency sampling according to the detection result of the collision detection device, and uploads the high-frequency sampled data to the collision detection end;

[0032] The collision detection end includes a feature extraction and fusion module, a feature matching module and a collision assessment module;

[0033] The feature extraction and fusion module stores a deep learning autoencoder model algorithm, a statistical algorithm, a linear transformation algorithm, and a PCA principal component analysis algorithm, and is used to parse and extract high-frequency sampling data uploaded by the device, obtain multi-dimensional features, and fuse the multi-dimensional features;

[0034] The feature matching module is used to match the feature fusion data with the data in the historical feature database and calculate the similarity distance;

[0035] The historical feature database stores abnormal data of a preset time period, and is offline at a data end, and the data end is connected to the collision detection end.

[0036] The beneficial effects of the present invention are as follows:

[0037] 1. Using deep learning autoencoding for feature extraction and introducing representation learning can solve the strong assumptions of single feature engineering on data distribution, make up for the defects of human prior features on large amounts of complex data, and effectively optimize the feature representation of high-frequency time series data; by integrating the features of three dimensions, namely linear transformation, statistics, and deep learning autoencoding, the problem of single feature extraction dimension is solved, the recall rate of collision accidents is effectively improved, and the accuracy of accident volume detection is improved.

[0038] 2. Introduce global abnormal data distribution information based on the historical feature data of a single device, and characterize the probability of collision by calculating the similarity distance with the historical data, so as to avoid misjudgment caused by a single local abnormal condition and enhance the detection of minor accidents, including vehicle scratches, chassis scrapes, etc., thereby improving the accuracy of the model and reducing the occurrence of false alarms. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic flow chart of the method of the present invention;

[0040] Figure 2 It is a schematic diagram of the multi-dimensional feature processing flow structure of the present invention;

[0041] Figure 3 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0042] The following description clearly and completely describes the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Example 1

[0044] Embodiment 1 of the present invention discloses a vehicle collision detection method based on feature time domain matching, such as Figure 1 As shown, the specific steps are as follows:

[0045] S1: Adopts adaptive data sampling algorithm to obtain high-frequency data packets uploaded in real time;

[0046] The sensors installed on the vehicle collect data according to the set algorithm;

[0047] In this embodiment, the vehicle driving data is collected by a six-axis gyroscope sensor, and the sampling device or the collision detection device installed on the vehicle is detected in real time to see whether a collision signal is output; if no collision signal is output, low-frequency sampling is performed, and if a collision signal is output, high-frequency sampling is performed. Vehicle collision is a low-probability event. When no collision is detected, that is, when the sensor device or the collision detection device does not output a collision signal, low-frequency sampling is performed according to preset parameters, and the point data of the low-frequency sampling is uploaded. If a collision is detected, that is, a large acceleration occurs or a vibration is detected, high-frequency data sampling is performed.

[0048] S2: Perform local anomaly detection based on the high-frequency data packets to determine whether there are local anomalies. If there are local anomalies, perform multi-dimensional feature processing on the data to obtain a feature set, otherwise terminate.

[0049] After obtaining the real-time high-frequency data packet, the sensor data of each high-frequency sampling point in the predetermined time-length trajectory segment is analyzed, and the analysis results of specific dimensions are vectorized respectively;

[0050] The vectorization processing, for example, each data point has acceleration values ​​of the x, y, and z axes (acc_x, acc_y, acc_z). After high-frequency sampling, each high-frequency package has 350 data points. Vectorization processing is performed on each dimension to obtain: vector value of the x-axis (acc_x_1, acc_x_2, …, acc_x_350), vector value of the y-axis (acc_y_1, acc_y_2, …, acc_y_350), vector value of the z-axis (acc_z_1, acc_z_2, …, acc_z_350).

[0051] The local anomaly detection may adopt methods such as threshold rule, isolation forest, LOF outlier detection, etc.

[0052] The multi-dimensional feature processing includes a first preprocessing and a second preprocessing, wherein the first preprocessing adopts different feature extraction methods to obtain multi-dimensional features, and the second preprocessing performs fusion and dimensionality reduction operations on the multi-dimensional features output by the first preprocessing to construct an output feature set;

[0053] like Figure 2As shown, in this embodiment, the feature extraction method of the first preprocessing includes deep learning autoencoding, statistical methods, and linear transformation; the statistical method can adopt conventional statistical calculation methods such as mean, variance, peak, first-order difference standard deviation, and maximum slope; the linear transformation adopts wavelet transform to decompose the high-frequency time series signal into different scales, and extract modulus maximum features, entropy features, energy features, Haar wavelet basis features, and symlets wavelet basis features.

[0054] The second preprocessing uses the RCA principal component analysis algorithm to fuse the different types of features obtained by the first preprocessing to obtain a feature set.

[0055] S3: Based on the feature set, feature matching is performed with data in a pre-built historical feature database, similarity distance is calculated, and the historical data with the highest similarity in the historical feature database is obtained;

[0056] The historical feature database is constructed as follows:

[0057] Perform local anomaly detection on the equipment belonging to each vehicle to obtain relatively abnormal high-frequency data packets within a preset time period;

[0058] The obtained high-frequency data packets are subjected to feature processing, and a feature set of each abnormal high-frequency data packet is extracted and stored to obtain a historical feature database.

[0059] The similarity distance is calculated as follows:

[0060] Normalize the data of the feature set to obtain the input data: X(1, 2, ..., x k ), k-dimensional normalized feature data;

[0061] Assume that there are N historical abnormal data in the historical feature database, namely: (1, 2, ..., U N ), where U 1 =(u 1 ,u 2 ,…,u d );

[0062] Match and calculate N data respectively, and the similarity distance D i The calculation formula is as follows:

[0063]

[0064] After that, get the minimum similarity distance D min =min(D 1 , D 2 , …, D N ). i，jRepresents the jth eigenvalue of the i-th historical abnormal data.

[0065] In this real-time example, when performing calculations on a computer, the above calculation process is optimized by constructing a KD tree, thereby reducing the consumption of online resources.

[0066] S4: inputting the minimum similarity distance obtained by calculating and matching with the historical feature database into the collision assessment model to output the collision detection result;

[0067] The specific process is as follows:

[0068] Calculate the minimum similarity distance D of the data point min The ranking from large to small in the historical anomaly database and the quantile rank of the ranking q , and the average speed avg_speed of the vehicle in 3 minutes, calculate the collision detection score, which is calculated as follows:

[0069]

[0070]

[0071] Among them, size is the evaluation parameter; the larger the score is, the more abnormal the data point is, that is, the greater the probability of collision.

[0072] Example 2

[0073] Embodiment 2 of the present invention discloses a vehicle collision detection system based on feature time domain matching, such as Figure 3 As shown, the system comprises:

[0074] At least one device end, a collision detection end and a historical feature database, wherein the device end is data-connected to the collision detection end and uploads data packets collected by the device end in real time, and the device end is provided with at least one collision detection device, and performs high-frequency sampling or low-frequency sampling according to the detection result of the collision detection device, and uploads the high-frequency sampled data to the collision detection end;

[0075] The collision detection end includes a feature extraction and fusion module, a feature matching module and a collision assessment module;

[0076] The feature extraction and fusion module stores a deep learning autoencoder model algorithm, a statistical algorithm, a linear transformation algorithm, and a PCA principal component analysis algorithm, and is used to parse and extract high-frequency sampling data uploaded by the device, obtain multi-dimensional features, and fuse the multi-dimensional features;

[0077] The feature matching module is used to match the feature fusion data with the data in the historical feature database and calculate the similarity distance;

[0078] The historical feature database stores abnormal data of a preset time period, and is offline at a data end, and the data end is connected to the collision detection end.

[0079] The present invention is not limited to the above-mentioned specific embodiments, but extends to any new features or any new combination disclosed in this specification, as well as any new method or process steps or any new combination disclosed.

Claims

1. A vehicle collision detection method based on feature time domain matching, It is characterized in that include: S1: Adopts adaptive data sampling algorithm to obtain high-frequency data packets uploaded in real time; S2: Perform local anomaly detection based on the high-frequency data packets to determine whether there are local anomalies. If there are local anomalies, perform multi-dimensional feature processing on the data to obtain a feature set, otherwise terminate. The multi-dimensional feature processing includes a first preprocessing and a second preprocessing, wherein the first preprocessing adopts different feature extraction methods to obtain multi-dimensional features, and the feature extraction methods of the first preprocessing include deep learning autoencoding, statistical methods, and linear transformation; the second preprocessing performs fusion and dimensionality reduction operations on the multi-dimensional features output by the first preprocessing to construct an output feature set, and the second preprocessing adopts a PCA principal component analysis algorithm to perform feature fusion on different types of features obtained by the first preprocessing; S3: Based on the feature set, feature matching is performed with data in a pre-built historical feature database, similarity distance is calculated, and the historical data with the highest similarity in the historical feature database is obtained; S4: Inputting the minimum similarity distance obtained by calculating and matching with the historical feature database into the collision assessment model to output the collision detection result.

2. The vehicle collision detection method according to claim 1, It is characterized in that The historical feature database is constructed as follows: Perform local anomaly detection on the equipment belonging to each vehicle to obtain relatively abnormal high-frequency data packets within a preset time period; The obtained high-frequency data packets are subjected to feature processing, and a feature set of each abnormal high-frequency data packet is extracted and stored to obtain a historical feature database.

3. The vehicle collision detection method according to claim 1, It is characterized in that The adaptive data sampling algorithm is specifically as follows: Real-time detection of whether the sampling device or the vehicle collision detection device outputs a collision signal; If no collision signal is output, low-frequency sampling is performed; if a collision signal is output, high-frequency sampling is performed.

4. The vehicle collision detection method according to claim 1, It is characterized in that In step S4, the collision detection result is specifically calculated as follows: S401: Obtain the rank and quantile of the minimum similarity distance in the historical anomaly database from large to small , and the average speed avg_speed of the vehicle during the preset time period; S402: The ranking rank and the corresponding quantile , calculate the evaluation parameter size and calculate the collision detection score. The calculation formula is as follows: in, represents the minimum similarity distance, a is the set ratio parameter, and its value is 10.

5. The vehicle collision detection method according to any one of claims 1 to 4, It is characterized in that In step S2, after obtaining the real-time high-frequency data packet, the sensor data of each high-frequency sampling point in the trajectory segment of the predetermined duration is analyzed, and the analysis results are vectorized respectively.

6. A vehicle collision detection system based on feature time domain matching, It is characterized in that The device comprises at least one device end, a collision detection end and a historical feature database. The device end is data-connected with the collision detection end, and data packets collected by the device end are uploaded in real time. The device end is provided with at least one collision detection device, and high-frequency sampling or low-frequency sampling is performed according to the detection result of the collision detection device, and the high-frequency sampling data is uploaded to the collision detection end; The collision detection end includes a feature extraction and fusion module, a feature matching module and a collision assessment module; The feature extraction and fusion module stores a deep learning autoencoder model algorithm, a statistical algorithm, a linear transformation algorithm, and a PCA principal component analysis algorithm, and is used to parse and extract high-frequency sampling data uploaded by the device end, obtain multi-dimensional features, and fuse the multi-dimensional features. The acquisition and fusion of the multi-dimensional features include a first preprocessing and a second preprocessing. The first preprocessing uses different feature extraction methods to obtain multi-dimensional features. The feature extraction method of the first preprocessing includes deep learning autoencoder, statistical method, and linear transformation. The second preprocessing fuses and reduces the multi-dimensional features output by the first preprocessing to construct an output feature set. The second preprocessing uses the PCA principal component analysis algorithm to fuse the different types of features obtained by the first preprocessing. The feature matching module is used to match the feature fusion data with the data in the historical feature database and calculate the similarity distance; The historical feature database stores abnormal data in a preset time period.

Images