IO-VNBD: Inertial and Odometry benchmark dataset for ground vehicle positioning

Low-cost Inertial Navigation Sensors (INS) can be exploited for a reliable solution for tracking autonomous vehicles in the absence of GPS signals. However, position errors grow exponentially over time due to noises in the sensor measurements. The lack of a public and robust benchmark dataset has however hindered the advancement in the research, comparison and adoption of recent machine learning techniques such as deep learning techniques to learn the error in the INS for a more accurate positioning of the vehicle. In order to facilitate the benchmarking, fast development and evaluation of positioning algorithms, we therefore present the first of its kind large-scale and information-rich inertial and odometry focused public dataset called IO-VNBD (Inertial Odometry Vehicle Navigation Benchmark Dataset). The vehicle tracking dataset was recorded using a research vehicle equipped with ego-motion sensors on public roads in the United Kingdom, Nigeria, and France. The sensors include a GPS receiver, inertial navigation sensors, wheel-speed sensors amongst other sensors found in the car, as well as the inertial navigation sensors and GPS receiver in an Android smart phone sampling at 10 Hz. A diverse number of driving scenarios were captured such as traffic congestion, round-abouts, hard-braking, etc. on different road types (e.g. country roads, motorways, etc.) and with varying driving patterns. The dataset consists of a total driving time of about 40 h over 1,300 km for the vehicle extracted data and about 58 h over 4,400 km for the smartphone recorded data. We hope that this dataset will prove valuable in furthering research on the correlation between vehicle dynamics and dependable positioning estimation based on vehicle ego-motion sensors, as well as other related studies.


a b s t r a c t
Low-cost Inertial Navigation Sensors (INS) can be exploited for a reliable solution for tracking autonomous vehicles in the absence of GPS signals. However, position errors grow exponentially over time due to noises in the sensor measurements. The lack of a public and robust benchmark dataset has however hindered the advancement in the research, comparison and adoption of recent machine learning techniques such as deep learning techniques to learn the error in the INS for a more accurate positioning of the vehicle. In order to facilitate the benchmarking, fast development and evaluation of positioning algorithms, we therefore present the first of its kind large-scale and information-rich inertial and odometry focused public dataset called IO-VNBD ( I nertial O dometry V ehicle N avigation B enchmark D ataset). The vehicle tracking dataset was recorded using a research vehicle equipped with ego-motion sensors on public roads in the United Kingdom, Nigeria, and France. The sensors include a GPS receiver, inertial navigation sensors, wheel-speed sensors amongst other sensors found in the car, as well as the inertial navigation sensors and GPS receiver in an Android smart phone sampling at 10 Hz. A diverse number of driving scenarios were captured such as traffic congestion, roundabouts, hard-braking, etc. on different road types (e.g. country roads, motorways, etc.) and with varying driving patterns. The dataset consists of a total driving time of about 40 h over 1,300 km for the vehicle extracted data and about 58 h over 4,400 km for the smartphone recorded data. We hope that this dataset will prove valuable in furthering research on the correlation between vehicle dynamics and dependable positioning estimation based on vehicle ego-motion sensors, as well as other related studies.

Data format Raw Parameters for data collection
The data was collected under a diverse number of environmental scenarios and vehicle motion states. The number of scenarios considered include bumps, hard braking, wet roads etc. See Table 4 for the full list of scenarios considered. Description of data collection The data was collected using four vehicles employing the sensors on a smartphone, GPS receiver and the sensors present in the sensor cluster of the vehicle. The smartphone data is sampled at 10 Hz with a GPS update rate of 1 Hz providing a total data size of about 2.2 million x 24, while the ECU recorded data is also sampled at 10 Hz with a total data shape of about

Value of the Data
• The dataset is large-scale and diverse, and it focuses on inertial vehicle navigation under complex environmental scenarios and vehicle motion states such as varying longitudinal accelerations, hard-brakes, yaw rates, velocities, mud roads, motorways, etc. (see Table 4 ). The dataset consists of measurements from a rich combination of ego-motion sensors such as accelerometers, gyroscope, magnetometers, wheel encoders, force sensors, etc.
• The data is useful to research institutions and industries in the benchmarking, fast development, evaluation and testing of vehicle positioning and tracking algorithms and techniques. • The data is useful for the robust training of supervised learning algorithms in learning the correlation between the dynamics of vehicles and their displacement, with applications in the tracking or positioning of vehicles and robots in GPS deprived environments using noisy low-cost sensors.

Data Description
The total dataset consists of about 100 h of recorded driving data on public roads by 8 different drivers with different driving styles as defined on Table 1 , where defensive driving refers to situations where the vehicle is turned at less than 0.3 g, swerved at less than 3.3 km/hr or decelerated at less than 0.3 g, whilst aggressive driving refers to respective situations above these thresholds [3] . The data is divided into sets based on cities and towns driven via, road conditions, weather conditions, driving length and time, driving style and driving features (see Tables A1-1 to A6 ). The dataset also contains more than 20 min of data recorded from the stationary vehicle to aid in the estimation of the sensors' bias. To add to the diversity of the data consisting of a number of complex driving scenarios as shown on Table 4 , the data was recorded with different tyre pressures. Datasets with each unique tyre pressures are indicated on Tables A1-1 to A5-2 using Table 2 as a guide. Tables A1-1 to A6 reveal more detailed information on each set of the data. The data logged from the vehicle's CAN bus are denoted with the prefix "V-" and the smartphone data denoted with the prefix "S-". The "S-" datasets are acquired from the sensors in a smartphone attached to the vehicle mimicking its motion. 1 While all the "V-" datasets were collected only in England, the "S-" datasets were collected in England, France and Nigeria.
Over the course of the data collection, communication difficulties between the GPS receiver and satellites were encountered. Information on data indexes recorded during these periods are provided in a file titled "GPS outages ". Where possible, the "S-" and "V-" datasets which were collected simultaneously, 2 are manually synchronised and stored in the folder named "Synchronised V and S datasets".
Importantly, despite the effort lent towards an accurate alignment of the smartphone's sensor axis with that of the vehicle, the precision of the measurements were interfered by vehicular vibrations averagely estimated to be about 0.15 g of acceleration and 0.08 rad/s of yaw rate particularly at peculiar scenarios such as hard brakes or over bumps. Information on the  amount of gravitational acceleration measured by each of the three axis are provided in the "S-" datasets to help in the correction of the measured acceleration. The data is stored in csv format at https://github.com/onyekpeu/IO-VNBD along with useful Python development tools.

Vehicle experiment setup
The vehicle used for the data collection exercise was a front wheel drive Ford Fiesta Titanium as shown in Fig. 2 . A Racelogic VBOX Video HD2 was used to record the data from the vehicle CAN bus as well as the corresponding GPS coordinates at each sampling instance. As shown in Figs. 1 and 2 , the GPS antenna was placed centrally at the top of the vehicle to ensure optimal signal reception. The Racelogic VBOX Video HD2 CAN -Bus data logger (10 Hz) was used to   Accelerator pedal position % activation record the data shown in Table 3 directly from the CAN bus of the vehicle with a sampling and update frequency of 10 Hz.

Smartphone measurement setup
A Ford Fiesta Titanium, Volvo XC70, Renault Mégane and Toyota Corolla Verso were used to collect the smartphone datasets. The smartphone was held with a phone holder attached to the vehicle as shown in Fig. 1 . Using the Androsensor app, all data were sampled every 0.1 s with a GPS (smartphone) update rate of 1 Hz. Figs. 1 and 2 show the axis alignment of the smartphone sensors. The smartphone sensors employed were a 3-axis accelerometer, a 3-axis gyroscope, a 3-axis magnetometer and heading, as well as the GPS latitude and longitude coordinates all present within the phone. Other information such as the vehicle's velocity and acceleration were recorded from the smartphone's GPS. Table 5 highlights the data recorded from the smartphone data. The datasets described in Tables A1-1 to A5-2 were collected using the Huawei P20 pro smartphone.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships which have, or could be perceived to have, influenced the work reported in this article.