A Review of Trajectory Datasets on Autonomous Driving

Abstract:
The advent of Autonomous Vehicles (AVs) heralds a transformative moment in transportation, with profound implications for urban landscapes and the potential for significant change human's life. The operationality of autonomous driving is divided into two systems: the perception system and the decision-making system. The former concerns itself with tasks such as vehicular positioning, mapping static impediments, and detecting and monitoring mobile obstacles. Meanwhile, the latter predominantly engages in tasks such as route planning and vehicular movement control.

   In the quest for advancing autonomous driving technology, the imperative for high-quality AV datasets to buttress algorithmic and modeling progress has crystallized. A plethora of autonomous driving datasets, including but not limited to BDD100K, Lyft Level 5 AV Dataset, and Waymo Open Dataset, have been promulgated this decade. Notably, the lion's share of these datasets directs its emphasis toward the perception system, utilizing camera data. Conspicuously absent is a substantive focus on trajectory data, such as GPS data, a crucial component for enhancing the decision-making system of AVs. This lacuna hampers the propulsion of research in AV-centric traffic flow analytics, especially at the microscopic level. Through scrupulous analysis of trajectory data, researchers can elucidate distinctive traffic flow phenomena, subsequently crafting decision models tailored to address these nuances. A salient application herein lies in the calibration of AV behavioral models, exemplified by IDM. Precise calibration ensures that these constructs faithfully mirror actual driving behaviors, thereby enriching insights gleaned from simulations. Another persistent challenge is the heterogeneity of data formats across datasets, placing an onerous task on researchers in data preprocessing. A clarion call for a universally accepted standard for trajectory datasets is manifestly evident.

   To fill these gaps, the present study embarks on a meticulous examination of trajectory datasets, amassed from vehicles equipped with adaptive cruise control. This endeavor encompasses systematic processing to streamline research utility. Our processing begins by standardizing these datasets into a uniform, user-friendly format, proposed as a prospective benchmark for ensuing trajectory datasets. Subsequently, we amplify the data's integrity through cleaning procedures and the excision of outliers. To buttress studies in car-following contexts, we extract car-following paired trajectories from the datasets and identify the primary information influencing the behavior of AV. Finally, we conclude with a comparative statistical analysis and calibration results across all datasets. All methodology, codes, and processed datasets underlying the study will be shared with the community after the acceptance of this paper. Thus, future developments can seamlessly integrate into our established framework, facilitating equitable and objective comparisons.