takeaway

Autonavi positioning business includes cloud positioning and terminal positioning two modules. Cloud location mainly solves Wifi fingerprint database, AGPS location, track mining and clustering, etc. On-end positioning solves the real-time positioning problems of mobile terminals and vehicle terminals. In recent years, with the development of positioning business, users have put forward higher requirements for positioning accuracy in urban canyons (high-rise buildings, viaducts, etc.).

In particular, because the positioning equipment is installed on the vehicle, on the one hand, it can carry more positioning sensors to solve the problems of special scenes. On the other hand, each sensor is connected to each other, which is conducive to high-precision algorithm design. These two points provide the possibility to further improve the navigation precision of vehicle and engine.

Urban canyon has always been a pain point for vehicle and engine end positioning. The reason is that the environment of urban canyon makes it impossible for users to receive GPS signals or GPS signals are interfered, leading to no GPS positioning results or poor positioning accuracy. This is the inherent shortcoming of “active location”, which cannot be overcome by algorithm.

To solve this problem, GPS+IMU multi-sensor fusion scheme is attracting more and more attention, because “passive positioning” IMU can make up for the shortcomings of GPS. In addition, the vehicle can also be equipped with odometer, vision equipment to form a richer multi-sensor fusion scheme.

For Autonavi, map data is the soul of the location business. Multi-sensor fusion is only a part of positioning business, how to combine multi-sensor and map data is always a problem we are thinking about.

Aiming at the vehicle-machine application, we use GPS, IMU, odometer and other sensors, combined with the advantages of Autonavi Map, put forward a multi-sensor fusion algorithm combined with Map Matching — GPS/IMU/MM fusion (software + hardware solution).

This paper summarizes the background of the vehicle multi-sensor fusion positioning project, which is established to provide users with good navigation positioning services. In order to solve the three major pain points reported by users: yaw recalcalculation, inability to locate and road catching errors, a software + hardware solution was proposed combining algorithm and data. Finally, the measured data is used to verify the improvement degree of user feedback.

Pain points for in-car apps

  • Yaw recalculation: refers to the position point drift caused by signal occlusion in elevated or urban canyon;
  • Unable to locate: it refers to the low accuracy of the calculation in the no-signal area (parking lot, tunnel), resulting in a large exit error;
  • Road catching error: it refers to the road catching error of main and secondary roads and overhead roads.

Among them, the direct cause of yaw recalculation and unpositioning is the poor positioning accuracy of GPS and DR dead reckoning. GPS positioning accuracy is determined by the observation environment and is difficult to improve; The accuracy of DR dead reckoning is related to the performance of DR algorithm, especially the calibration accuracy of odometer system error and gyro zero deviation. The direct cause of road catching error is that the correct road and the wrong road are too close to each other, so they cannot be distinguished due to the limitation of positioning accuracy. The root cause is that only location information is used for road capture, without giving play to the value of other data.

Relative term

GlobalPositioningSystem (GPS) : the GlobalPositioningSystem developed by the us department of defense. By receiving GPS signals, the user equipment can obtain the distance observation value of the user equipment and the satellite, and obtain the three-dimensional coordinate and heading information of the user equipment through special algorithm processing. Using different types of observations and algorithms, the positioning accuracy ranges from centimeter to 10 meters. The advantage of GPS is high precision, error does not diverge with time, the disadvantage is that the requirements of visibility, positioning range can not cover the indoor.

IMU(Inertial measurementunit) : Inertial measurementunit. Including gyroscopes and accelerometers. Gyroscope is used to measure angular velocity of three axes of the object and calculate the attitude of the carrier. The accelerometer measures the linear acceleration of the three axes of the object and can be used to calculate the velocity and position of the carrier. The advantage of IMU is that it does not require visibility and the positioning range is the whole scene. The disadvantage is that the positioning accuracy is not high, and the error diverges with time. GPS and IMU are two complementary positioning technologies.

MM(Map matching) : Map matching. The technology combines user location information with map data to calculate the user’s location on the road in the map data.

Technical solution

The vehicle-vehicle integration positioning project solves the road-level positioning problem. Limited by hardware performance, there are two common technical solutions on the market, as shown in Table 1 below:

These two technical schemes involve three technical means, which have their own advantages and complement each other in scene coverage and accuracy. As shown in Table 2:

Table 1 shows that the existing general solutions on the market cannot completely solve the three problems of yaw recalculation, unpositioning and road grasping error, especially the road grasping error. Therefore, at the technical level, we integrate the two sets of general schemes and propose a set of soft + hard (GNSS+MM+DR) scheme. At the algorithm level, relying on the data advantages of Autonavi and taking the data fusion module as the core, on the one hand, it improves the reliability of positioning results and makes up for the shortage of hardware performance, on the other hand, it designs a special algorithm for the problem of road capture error.

Furthermore, the three problems reported by users are deconstructed into three algorithmically solved problems: device error calibration, scene identification and data fusion. As shown in the figure:

Function module

Car-machine fusion positioning includes DataAdaptive Layer, Aided NavigationLayer and NavigationLayer. The data adaptation layer is responsible for standardizing different inputs and synchronizing signals. The support layer of the algorithm computes the intermediate results to serve the fusion layer. As the core of the whole system, the fusion layer is responsible for fusing the data output by the algorithm support layer to obtain reliable navigation information. Figure 3 lists the location of each layer and the specific functional modules of each layer:

Below, the functional module is divided into two parts: basic module and characteristic function.

Basic module

GPS quality assessment

The function of GPS quality assessment module is to calculate GPS position, speed, heading Angle and global reliability index. Reliability indicators are projected into the state space (GOOD, DOUBT, BAD, ABNORMAL) according to their size, and the value of the state space represents the quality of GPS data. As shown in the figure:

GPS quality evaluation has two purposes: first, to determine whether to use GPS data for device error calibration or some state judgment (such as turning behavior, static state, etc.); Second, in the data fusion module, it provides reference for setting the variance-covariance matrix of GPS observations.

The compensation device

In the absence of GPS signal, location can only rely on DR algorithm. The accuracy of DR algorithm mainly depends on the errors of IMU (gyroscope and accelerometer) and velocimeter. The gyroscope error will cause the quadratic increase of position error with time, and the velocimeter error will cause the linear increase of position error with time, as shown in Figure 5:

In order to improve the positioning accuracy without GPS signal, device error compensation must be carried out.

The main function of the compensation module is to use GPS data to compensate the error parameters of velocity sensor (scale factor) and IMU (gyroscope celestial scale factor and gyroscope three-axis zero deviation). The purpose of compensation is to obtain reliable navigation information only by DR algorithm in the scenario with no GPS signal or weak GPS signal.

DR algorithm

The DR algorithm will calculate the navigation state (state, speed and position) from the sensor observations to the next time. DR algorithm includes posture arrangement and position arrangement.

Attitude layout uses AHRS (Attitude andheading reference system) fusion algorithm, and the Attitude information of vehicle and vehicle is output after processing. The pose choreography process is shown in Figure 6:

Position arrangement refers to integrating the observation value of speedometer to get the position of vehicle and machine based on the result of attitude arrangement.

Fusion algorithm

The vehicle-vehicle fusion positioning project uses GNSS, MM and DR technologies, each of which can provide all or part of vehicle-vehicle navigation information, as shown in Table 3. Take location information as an example, GNSS, MM and DR all output the position of vehicle and machine, but because different technical means have their own error sources, the positioning results of different technologies are different.

Therefore, the fusion algorithm has two purposes: first, to fuse the navigation information of different technologies into a unique navigation information, so that its reliability is higher than before the fusion; Second, device errors (gyroscope zero deviation, velocimeter scale error and navigation error, etc.) are estimated.

The fusion algorithm is based on Kalman filter, and the key lies in model building and model parameter setting. Kalman filter model is composed of state transition equation and observation equation. The state transition equation represents the transition relation between adjacent navigation states, which is realized by constructing navigation error differential equation. Model parameters refer to state transition noise and observation noise, and the setting of observation noise is related to GPS quality assessment module.

After Kalman filtering, the optimal estimation of navigation error is obtained, as shown in FIG. 7. That is, the optimal estimation of navigation information is obtained after compensation.

features

Identification of main and auxiliary roads

In the past, the main and secondary routes were identified based on the relationship between the position and direction of the DR output and the two candidate routes, and the closest candidate route was selected as the output. However, for the road-level positioning system, the DR output error is the same magnitude as the difference between the two roads, and the probability of misjudgment is high. Therefore, it is necessary to solve this problem from some driving characteristics, such as turning and lane change near the intersection.

As shown in the figure, specific steps are as follows:

  • The characteristics of driving behavior were extracted and the transfer probability of feature information was calculated.
  • The convolution sum is calculated according to the classification of DR accuracy and the final probability is calculated.

Elevated recognition

In the past, viaduct identification strategy was judged by elevation integration and threshold method, and the identification effect was affected by slope Angle error and velocity error. Among them, the velocity error is directly proportional to the elevation integration error, which is the main reason affecting the accuracy of elevated identification. To overcome this shortcoming, we combined MM technology to calculate the proximity between road slope and output pitch Angle, so as to avoid introducing speed error. The elevated identification process is shown in the figure:

Parking lot identification

Parking lot identification is a new module, parking lot positioning and navigation is the pre-work. The main purpose of parking lot positioning navigation is to navigate the user to the designated parking space, which involves a series of problems such as indoor and outdoor scene map switching, floor to floor map switching and navigation. The purpose of parking lot recognition is to support this map switching.

Parking lot includes entry and exit parking lot recognition and cross – layer recognition. Entry and exit parking lot identification refers to the use of parking lot no GPS signal, up and down slope, low speed, elevation changes and a series of characteristics to judge whether vehicles and vehicles enter and exit parking lot. Cross-storey identification of parking lot refers to the use of the characteristics of up-slope and down-slope, elevation change and so on to judge whether there is cross-storey behavior in the parking lot. The identification process is shown in the figure:

The effect

To verify the effect of the algorithm in this project, we collected measured data and verified it from the following two aspects:

  • Verify the improvement degree of the algorithm to the user’s pain points (yaw recalcalculation, unable to locate, and);
  • Comparison with rival products and Autonavi mobile phone positioning terminal product performance.

Fusion of bleaching

Aiming at the yaw recalculation (position drift) problem of viaduct and urban canyon, the vehicle-machine algorithm is improved in the following two aspects:

  • Multivariate reference: combine motion trend, sensor information and map data to normalize GPS reliability;
  • Scene classification: Judge whether there is occlusion by referring to road attributes on the map and GPS signal distribution.

Two rounds of data were collected under the viaduct and processed by vehicle and machine software and a similar software in the market. The results are shown in Figure 14. According to the test of nearly half a year, the anti-drift ability of this project is obviously better than that of the traditional scheme in the case of GPS occlusion.

The device calibration

In order to verify the influence of gyro dynamic zero bias estimation on DR direction and position calculation accuracy, data of wangjing SOHO parking lot are collected in this project. After calculation, the results are as follows. The test results show that the DR accuracy is significantly improved after dynamic zero offset compensation:

  • Zero bias: dynamic zero bias estimation ensures that the gyroscope error magnitude is 0.01 degree /s;
  • Direction: the direction error of the exit of the parking lot is reduced to less than 40%, and the direction accuracy is increased by more than 2 times;
  • Location: the location error of the exit of the parking lot is reduced to less than 25%, and the location accuracy is increased by more than 4 times.

Identification of main and auxiliary roads

In order to calculate the success rate of the recognition of main and auxiliary roads, the recognition effect of nearly a thousand main and auxiliary roads is counted, and the recognition rate is more than 90%, which is greater than 75% of the product of a certain manufacturer.

Horizontal/vertical comparison

Finally, we made a horizontal comparison with the middle and high-end competitive products in the market and a vertical comparison with Autonavi mobile positioning products. The horizontal comparison results show that the position error, direction error and velocity error that do not exceed a certain threshold account for more than 90% when the device cost is less than 10% of the cost of the competing product, which is increased by 1%-5% compared with the competing product. The recognition accuracy of main auxiliary path is better than 90%, which is 15% higher than that of competing products.

The longitudinal comparison results show that under different scenarios (viaduct, urban canyon, roundabout, parking exit, etc.), the proportion of location error that does not exceed a certain threshold increases by 15%~60%, which is because the vehicle-machine algorithm specially designs and optimizes the algorithm for special scenarios (no GPS or weak GPS scenarios). The proportion of position error in all scenarios increased by about 20%.

summary

Aiming at the three pain points proposed by users, this paper proposes a set of vehicle-mounted multi-sensor fusion positioning scheme combining multi-sensor fusion and map matching, and applies it to practice, which improves the positioning accuracy in urban canyon and achieves good results.

However, the positioning accuracy of urban canyon is difficult to be solved completely, and it seems to be a problem with no end. To this end, from the perspective of users, we need to keep thinking about what sensor technology we need, what algorithm we should design, and how to extract the maximum value of data.