Authors
Shree Reddy, Shreya Gupta, Ryan Morokutti, Alan Wong, Sanjay Lall, Nick Landolfi
Abstract
In our increasingly fast-paced and technologically-driven society, autonomous vehicles are the future of smart and sustainable living. To create a basis for vehicles with a comfortable and safe user experience, our research centers on establishing the fundamental differences between human driving preferences and those of autonomous vehicles. Specifically, this paper aims to identify the differences, if any, in the execution of maneuvers by autonomous vehicles and in the linearity of their trajectories. Through our research, we analyze two comprehensive data sets, one featuring GPS coordinates and camera footage of human-driven cars at intervals of 100 milliseconds, the other featuring an autonomous vehicle’s position at every 200 milliseconds. We first use coordinate differencing and spline interpolation to compute a reasonable approximation for jerk, then linear regression to examine lane-keeping behavior. The statistical significance in relation to both is evaluated in order to determine the reliability of the data. In relation to the measurement of jerk, these statistical tests revealed, along with our algorithmic analysis, that human-driven cars make sharper turns and drive less smoothly than autonomous cars- challenging our original hypotheses. However, neither human-driven nor autonomous vehicles were shown to have statistically significant variance in the linearity of their trajectories. These results will provide the basis for further research as to the implications of these findings, and their use in future technologies for the eventual creation of safe and comfortable autonomous systems.
1 Background
In the last 10 years, autonomous vehicles have advanced significantly in terms of safety, viability, and navigation technology [7,13]. This technology has the potential to increase mobility by reducing traffic congestion and transportation costs. By decreasing emissions and energy consumption, autonomous vehicles could reduce the damage done to our environment [4,11]. They also have the potential to make commuting far more safe by eliminating human error. Urban landscapes will shift to accommodate the potential increase in commuters, as will many aspects of urban and suburban life.
In order for these changes to happen, establishing the fundamental differences between AV and human driving is critical to the creation of a truly viable autonomous system. In order to coexist on roads safely, both humans and autonomous vehicles have to predict each other’s actions, which becomes impossible without knowing the differences between each other’s driving patterns [2,8]. Furthermore, enjoyment of initial user experience is key to gaining the trust of both passengers and those sharing the road with autonomous vehicles [1,6]. Previous research has shown that humans are more comfortable with what they are familiar with [9]. If autonomous vehicles drive significantly differently from human-driven cars, people are less likely to adopt them. The benefits of autonomous vehicles will never be realized unless they are adopted, and user experience is crucial in facilitating this. Research into the fundamental differences between human and autonomous vehicles will allow for a more educated discussion, and eventually, implementation of these ideas into a tangible solution.
The goal with our research is to establish the fundamental differences between human and autonomous driving. The safety and viability of autonomous vehicles in current society is extremely important in terms of public trust and support of this advancement [1,6]. For this reason, preliminary research into the fundamental differences between human and autonomous vehicles is necessary for the eventual realization of a sustainable future by way of autonomous driving.
2 Research Question
What are the fundamental differences between the driving patterns of human-driven and autonomous vehicles?
3 Hypotheses
- Autonomous cars are less adept at driving smoothly, which we measure as maintaining a constant rate of acceleration.
- Autonomous drivers make sharper turns than human drivers do.
- Autonomous cars are likely to have a more linear trajectory when driving straight.
4 Methods and Materials
We conducted our research using the open source Lyft Level 5 Prediction Dataset [3] and the Interaction Dataset [14]. The Lyft Level 5 Prediction Dataset contains scenes capturing the position of one of it’s fleet autonomous vehicles and the surrounding environment and the Interaction Dataset contains data about the trajectories of human-driven cars. We manually categorized and cut up the scenes in both datasets to isolate when the car changed speed, made a turn, or drove in a straight lane.
4.1 Smooth Driving
We quantified how smoothly the vehicles drove by finding the mean jerk during a period where it accelerated or decelerated. For both human-driven and autonomous vehicles we used the data describing the location of the vehicle at a specific time. We approximated the velocity of the vehicle at a given time by finding the slope between the point that gives the location of the vehicle at the time we want its velocity and the point after it. We repeated this two more times to approximate the jerk of the vehicle and then used spline interpolation to fit a function to the data points. We found the average value of the function and got our final result, the average jerk of the vehicle throughout its entire trajectory.
4.2 Smooth Turning
We quantified how smoothly the vehicles turned by finding the mean rotational jerk during a period where it accelerated or decelerated. For both human-driven and autonomous vehicles we used the data describing the location of the vehicle at a specific time. We used the same method to find the velocity of the vehicle as we used in the above section. We then found the centripetal acceleration by using the formula for centripetal acceleration: acceleration = velocity ** 2 / radius of curvature. At this point we had the velocity and time, but we didn’t have the radius of curvature. Because the car’s trajectory isn’t a perfect circle, the radius of curvature was constantly changing. We approximated the radius of curvature at each point by fitting a unique circle to the current point, the point before it, and the point after it and then calculating the radius of the circle. Once we had the centripetal acceleration, we approximated the jerk using the same method of approximating the derivative as the section above and then found the average jerk using the same method as the section above.
4.3 Linearity of the Trajectory
For all of the scenes which included the vehicle driving on a straight lane, we used the data given by the datasets about the location of the vehicle at a specific time . We then used a linear regression to fit a line to the data points of the trajectory of the vehicles and then found how much the actual trajectory varies from the fitted line. We found the coefficient of correlation to determine how precisely the data points fit the line which would tell us how straight the trajectory of the car was.
5 Results
5.1 Smooth Driving
We found that on average human-driven vehicles have more jerk (0.516 m/s3) when changing speeds than autonomous vehicles (0.230 m/s3 ). We ran the data through an independent-samples t-test to determine if the results were significant, which they were (p <0.05). This challenges our hypothesis that autonomous vehicles drive less smoothly than humans.
| Average Jerk | n | mean | Std. Dev. | Std. Error Mean |
| Human Data | 16 | 0.516 m/s3 | 0.366 | 0.0915 |
| Autonomous Vehicle Data | 44 | 0.230 m/s3 | 0.193 | 0.0291 |
| 95% Confidence Interval of the Difference (Lower) | 95% Confidence Interval of the Difference (Upper) | T | df | Sig. (2-tailed) |
| 0.899 | 0.4821 | 2.9787 | 33.1059 | 0.0058 |
In order to adequately interpret these figures for jerk, our reader should know first and foremost that humans’ threshold for jerk discomfort lies at about 0.9 meters per cubic second, although in certain situations it is estimated as low as 0.5 meters per cubic second [10]. For some professions, not limited to self-driving car engineers, jerk arises on a daily basis. Railroads are typically designed with up to 0.35 meters per cubic second of jerk as the goal, but engineers allow for a maximum of 0.5 meters per cubic second in the case that there are sharp turns to be made. Although humans report discomfort when the elevator they are standing increases its acceleration at a rate of 2.0 meters per cubic second or higher, hospital elevators set their maximum at 0.7 meters per cubic second. If nothing else, these outside examples reveal the plausibility of the numbers we provide in this paper.
The above graphs are samples of an individual vehicle from the human-driven dataset (left) and an individual vehicle from the autonomous dataset (right). These are not representative of the data as a whole.
5.2 Smooth Turning
We found that on average human-driven vehicles have more jerk (0.645 m/s3) when turning than autonomous vehicles (0.484 m/s3 ). We ran the data through an independent-samples t-test to determine if the results were significant, which they were (p <0.05). The data challenges our hypothesis that autonomous vehicles take sharper turns because the average jerk for the autonomous vehicles actually comes out as the lower of the two. A sharp turn would mean a more sudden change in both linear and centripetal acceleration. While this is not necessarily equated to larger jerk values, in practice these tend to be correlated.
| Coefficient of Correlation | n | mean | standard deviation | standard error of the mean |
| Human Data | 89 | 0.645 m/s3 | 0.242 | 0.0257 |
| Autonomous Vehicle Data | 60 | 0.484 m/s3 | 0.457 | 0.0590 |
| 95% Confidence Interval of the Difference (Lower) | 95% Confidence Interval of the Difference (Upper) | t | df | Sig. (2-tailed) |
| 0.033 | 0.289 | 2.5026 | 75.898 | 0.0247 |
The above graphs are samples of an individual vehicle from the human-driven dataset (left) and an individual vehicle from the autonomous dataset (right). These are not representative of the data as a whole.
5.3 Linearity of the Trajectory
We found that on average, autonomous vehicles were not proven to have a significantly more linear trajectory than self-driving vehicles. We ran the data through an independent-samples t-test to determine if the results were significant, which they were (p <0.05). This challenges our hypothesis that autonomous vehicles favor the center of the lane more than human drivers, as our hypothesis is now rather null.
| n | Mean correlation coefficient | standard deviation | standard error of the mean | |
| Human Data | 45 | 0.999 | 0.0002 | ≈0 |
| Autonomous Vehicle Data | 51 | 0.990 | 0.02295 | 0.003 |
| 95% Confidence Interval of the Difference (Lower) | 95% Confidence Interval of the Difference (Upper) | t | df | Sig. (2-tailed) |
| -0.0557 | 0.0737 | 0.2794 | 48.0937 | 0.7799 |
The above graphs are samples of an individual vehicle from the human-driven dataset (top) and an individual vehicle from the autonomous dataset (bottom). These are not representative of the data as a whole.
6 Limitations
One limitation is that the data for human-driven cars was taken in the US, Germany, and China while the autonomous driving data was only taken in the US. Drivers from different countries may have different driving patterns which could have skewed the data. Another limitation is that the vehicles in both the human and autonomous dataset drove on different roads. Differences in the shape and condition of the roads, such as lane width, could have changed driving patterns. A third limitation is that we didn’t have information about the environment for the human data. Certain environmental factors, such as the presence of traffic agents near the vehicle, could have influenced driving patterns as well.
Our calculations for the jerk come with their own set of limitations. The simplest way to calculate jerk from a discrete list of positions and timestamps, which was essentially what our data was, would be to do differencing. In our case, differencing just means to subtract each known coordinate of a vehicle’s trajectory from the preceding coordinate, divide by the time it takes the vehicle to travel that distance to get an average velocity over a time interval, and then repeat the process of subtracting and dividing two more times to get acceleration and then jerk. The problem with this procedure is that it assumes velocity is constant between two positions, which leaves acceleration as the jump between two discontinuous average velocities. If acceleration is then unreasonably high, the jerk can only be more so.
7 Future Research
In the future, research that takes into account the different driving patterns of different countries would be beneficial in determining which countries would be more likely to adopt autonomous vehicles. Adoption of autonomous vehicles in these countries would be the first step towards making autonomous vehicles more safe and a viable mode of transport for all.
Additionally, further research as to the variance in merging between autonomous and human-driven vehicles is another area that should be explored. This understanding will provide a basis for the implementation of various methods to integrate autonomous vehicles into real-life driving situations. In this case, anecdotal data is important too. Feedback from human passengers, for example, like whether they notice the difference in jerk and prefer one driving style over another, would be extremely useful. This data would drive research further into the comfort and practicality of autonomous vehicles in everyday life, which is key when designing what may very well become a consumer product.
As for jerk, to move past our algorithmic and calculation limitations, we would need to find a function which better approximates jerk to the point where we can accurately compare the jerk of human-driven versus autonomous vehicles. Once this is achieved, this visualization of the differences in jerk will need to be compared to the tangible effects of human passengers. Our plan is to collect our own data, which admittedly will have to be much smaller in scale but which will allow us to draw a more direct comparison. Like many other studies have done, we will find a driving course where we can run tests of human and autonomous vehicles one at a time [5]. Using the same route, perhaps a double-standard lane change course to comprehensively test vehicles’ steering control, will eliminate our main source of error- that the human data and the autonomous data were conducted in different circumstances and under different time intervals [12]. Conducting our own data collection would enable us to equip our cars with a jerk meter and a GPS, which would return actual measurements for jerk and lane-keeping. These future experiments and the results they will yield will bring us one step closer to charting the viability of autonomous vehicles in human society.
8 Acknowledgements
We would like to thank Prof. Sanjay Lall for his assistance and guidance in terms of helping us conduct proper research within the field of self-driving cars. We would also like to thank Nick Landolfi for being a mentor to our group and providing us not only with informative feedback and helping us find reliable data during this pandemic, but also for guiding us towards the right path whenever we experienced difficulty within a particular problem our project had for us. Lastly, we would like to thank Prof. Tsachy Weissman and Cindy Nguyen for creating and managing this internship in which we were able to get a better understanding of the fields we are passionate about.
9 References
[1] Choi, J. K. and Ji, Y. G. (2015) Investigating the Importance of Trust on Adopting an Autonomous Vehicle, International Journal of Human–Computer Interaction, 31:10, 692-702, DOI: 10.1080/10447318.2015.1070549
[2] Hancock, P. A., Nourbakhsh, I., and Stewart, J. (2019). On the future of transportation in an era of automated and autonomous vehicles. Proceedings of the National Academy of Sciences of the United States of America, 116(16), 7684–7691. https://www.pnas.org/content/116/16/7684
[3] Houston, J., Zuidhof, G., Bergamini, L., Ye, Y., Jain, A., Omari, S., Iglovikov, V., and Ondruska, P. (2020). One Thousand and One Hours: Self-driving Motion Prediction Dataset. Lyft Inc. https://level5.lyft.com/dataset/.
[4] Igliński, H. and Babiak, M. (2017, January 1) Analysis of the Potential of Autonomous Vehicles in Reducing the Emissions of Greenhouse Gases in Road Transport. ScienceDirect. https://www.sciencedirect.com/science/article/pii/S1877705817326073
[5] Kong, J., Pfeiffer, M., Schildbach, G., and Borelli, F. (2015) Kinematic and Dynamic Vehicle Models for Autonomous Driving Control Design. https://borrelli.me.berkeley.edu/pdfpub/IV_KinematicMPC_jason.pdf
[6] Lankton, N. K., McKnight, D. H., and Tripp, J. (2015) “Technology, Humanness, and Trust: Rethinking Trust in Technology,” Journal of the Association for Information Systems: Vol. 16 : Iss. 10 , Article 1.
[7] Moore, M. M. and Lu, B. (2011, June 2) Autonomous Vehicles for Personal Transport: A Technology Assessment. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1865047
[8] Oliveira, L., Proctor, K., Burns, C. G., & Birrell, S. (2019). Driving Style: How Should an Automated Vehicle Behave? Information, 10(6), 219. DOI:10.3390/info10060219
[9] Sibi, S., Baiters, S., Mok, B., Steiner, M., and Ju, W. (2017, June 11–14) Assessing driver cortical activity under varying levels of automation with functional near infrared spectroscopy. In Proceedings of the 2017 IEEE Intelligent Vehicles Symposium (IV), Redondo Beach, CA, USA, pp. 1509–1516.
[10] Svensson, L., & Eriksson, J.C. (2015). Tuning for Ride Quality in Autonomous Vehicle : Application to Linear Quadratic Path Planning Algorithm. http://www.diva-portal.org/smash/get/diva2:839140/FULLTEXT01.pdf
[11] Talebpour, A., Mahmassani, H. S., & Elfar, A. (2017). Investigating the Effects of Reserved Lanes for Autonomous Vehicles on Congestion and Travel Time Reliability. Transportation Research Record, 2622(1), 1–12.
[12] Tan, H. and Huang, J. (2012). Experimental Development of a New Target and Control Driver Steering Model Based On DLC Test Data. https://ieeexplore.ieee.org/document/6094217
[13] Thrun, S. (2010). Toward robotic cars. Communications of the ACM, 53(4), 99–106. https://dl.acm.org/doi/10.1145/1721654.1721679
[14] Zhan, W., Sun, L., Wang, D., Shi, H., Clausse, A., Naumann, M., Kummerle, J., Konigshof, H., Stiller, C., de La Fortelle, A., Tomizuka, M. (2019). INTERACTION Dataset: An INTERnational, Adversarial and Cooperative moTION Dataset in Interactive Driving Scenarios with Semantic Maps.https://interaction-dataset.com/.