Coupes and convertibles were the most expensive used car types in the United States as of February 2023, priced on average at around 49,800 and 46,700 U.S. dollars respectively. In contrast, used wagons and hatchbacks were more affordable, at an average of 20,000 and 24,200 U.S. dollars. The overall used vehicle average list price had been steadily rising between mid-year 2020 and mid-year 2022, but dipped in June 2023.

The U.S. auto industry sold nearly three million cars in 2024. That year, total car and light truck sales were approximately 15.9 million in the United States. U.S. vehicle sales peaked in 2016 at roughly 17.5 million units. Pandemic impact The COVID-19 pandemic deeply impacted the U.S. automotive market, accelerating the global automotive semiconductor shortage and leading to a drop in demand during the first months of 2020. However, as demand rebounded, new vehicle supply could not keep up with the market. U.S. inventory-to-sales ratio dropped to its lowest point in February 2022, as Russia's war on Ukraine lead to gasoline price hikes. During that same period, inflation also impacted new and used car prices, pricing many U.S. consumers out of a market with increasingly lower car stocks. Focus on fuel economy The U.S. auto industry had one of its worst years in 1982 when customers were beginning to feel the effects of the 1973 oil crisis and the energy crisis of 1979. Since light trucks would often be considered less fuel-efficient, cars accounted for about 77 percent of light vehicle sales back then. Thanks to improved fuel economy for light trucks and cheaper gas prices, this picture had completely changed in 2020. That year, prices for Brent oil dropped to just over 40 U.S. dollars per barrel. The decline occurred in tandem with lower gasoline prices, which came to about 2.17 U.S. dollars per gallon in 2020 - and cars only accounted for less than one-fourth of light vehicle sales that year. Four years on, prices are dropping again, after being the highest on record since 1990 in 2022.

United States Motor Vehicle Sales: Average Price data was reported at 13,105.000 USD in 2010. This records an increase from the previous number of 12,518.000 USD for 2009. United States Motor Vehicle Sales: Average Price data is updated yearly, averaging 12,098.000 USD from Dec 1990 (Median) to 2010, with 21 observations. The data reached an all-time high of 13,451.000 USD in 2007 and a record low of 8,691.000 USD in 1990. United States Motor Vehicle Sales: Average Price data remains active status in CEIC and is reported by Bureau of Transportation Statistics. The data is categorized under Global Database’s United States – Table US.RA004: Motor Vehicle Sales: Average Price.

In 2024, the auto industry in the United States sold approximately 15.9 million light vehicle units. This figure includes retail sales of about three million passenger cars and just under 12.9 million light trucks. Lower fuel consumption There are many kinds of light vehicles available in the United States. Light-duty vehicles are popular for their utility and improved fuel economy, making them an ideal choice for savvy consumers. As of Model Year 2023, the light vehicle manufacturer with the best overall miles per gallon was Kia, with one gallon of gas allowing for 30.4 miles on the road. Higher brand satisfaction When asked about light vehicle satisfaction, consumers in the United States were most satisfied with Toyota, Subaru, Tesla, and Mercedes-Benz models. Another survey conducted in 2018 and quizzing respondents on their stance regarding the leading car brands indicated that Lexus was among the most dependable brands based on the number of problems reported per 100 vehicles.

United States Motor Vehicle Sales: Average Price: Used Vehicle data was reported at 8,786.000 USD in 2010. This records an increase from the previous number of 8,483.000 USD for 2009. United States Motor Vehicle Sales: Average Price: Used Vehicle data is updated yearly, averaging 8,130.000 USD from Dec 1990 (Median) to 2010, with 21 observations. The data reached an all-time high of 8,786.000 USD in 2010 and a record low of 5,857.000 USD in 1990. United States Motor Vehicle Sales: Average Price: Used Vehicle data remains active status in CEIC and is reported by Bureau of Transportation Statistics. The data is categorized under Global Database’s United States – Table US.RA004: Motor Vehicle Sales: Average Price.

Used Car Prices YoY in the United States decreased to 0.10 percent in February from 0.80 percent in January of 2025. This dataset includes a chart with historical data for the United States Used Car Prices YoY.

Sales of used light vehicles in the United States came to around 38.9 million units in the third quarter of 2024. The same period, approximately 15.6 million new light trucks and automobiles were sold here. Declining availability of vehicles In the fourth quarter of 2023, about 288.5 million vehicles were in operation in the United States, an increase of under one percent year-over-year. The rising demand for vehicles paired with an overall price inflation lead to a rise in new vehicle prices. In contrast, used vehicle prices slightly decreased. E-commerce: a solution for the bumpy road ahead? Financial reports have revealed how the outbreak of the coronavirus pandemic has triggered a shift in vehicle-buying behavior. With many consumer goods and services now bought online due to COVID-19, the automobile industry has also started to digitally integrate its services online to reach consumers with a preference for contactless test driving amid the global crisis. Several dealers and automobile companies had already begun to tap into online car sales before the pandemic, some of them being Carvana and Tesla.

Number of units and total sales of new motor vehicles by vehicle type and origin of manufacture, monthly.

This dataset contains data from a survey of new-car buying households in 13 US states conducted December 2014 to January 2015. The original study is described in these technical reports:

Kurani, K S., N. Caperello, J. TyreeHageman New Car Buyers' Valuation of Zero-Emission Vehicles: California, Institute of Transportation Studies, University of California, Davis, Research Report UCD-ITS-RR-16-05 (2016). https://escholarship.org/uc/item/28v320rq

Kurani, K.S., N. Caperello, J. TyreeHageman NCST Research Report: Are We Hardwiring Gender Differences into the Market for Plug-in Electric Vehicles? Institute of Transportation Studies, University of California, Davis, Research Report UCD-ITS-RR-18-05 (2018). https://itspubs.ucdavis.edu/publication_detail.php?id=2888

This dataset is associated specifically with a subsequent technical report:

Kurani, K.S. and K. Buch Across Early Policy and Market Contexts Women and Men Show Similar Interest in Electric Vehicles, National Center for Sustainable Transportation, University of California, Davis, Research Report. 2019. https://escholarship.org/uc/item/9zz8n5x5

Data are from households who had a acquired at least one household vehicle as new (rather than used) since January 2008. The questionnaire was administered on-line to households in the following US states: California, Connecticut, Delaware, Maine, Maryland, Massachusetts, New Hampshire, New Jersey, New York, Oregon, Rhode, Island, Vermont, and Washington. Most of these states are so called "ZEV states," i.e., they had adopted California's Zero Emission Vehicle (ZEV) Mandate. Those states that were not ZEV states were included to facilitate regional analysis or because they were otherwise important to the initial launch of retail ZEV sales in 2011. The primary regional analysis was for the Northeast States for Coordinated Air Use Management (NESCAUM). The NESCAUM member states are Connecticut, Maine, Maryland, Massachusetts, New Hampshire, New Jersey, New York, Rhode Island, and Vermont. The total sample size is 5,654 for all states; individual state samples sizes are available in the above referenced, Kurani et al (2016).

Analyses were conducted at the state and regional, i.e., NESCAUM, levels. Thus, there are individual data sets for each state for which there is a state-level analysis (California, Delaware, Maryland, Massachusetts, New Jersey, New York, Oregon, and Washington) and NESCAUM. Data for California are included in this release despite the fact its analysis was previously conducted under a separate study. California serves as the reference case because it has the most supportive policy and market context for ZEVs and its analysis is specifically referenced in the report associated with these data sets.

Since the goal was to produce the best possible analysis for each state or region, there are differences in their data sets. While variable names and codes follow consistent rules across all the data sets, which variables are in the data does vary across states and the NESCAUM region. The data released here are those required to replicate the analyses in the associated report.

For each state and region, data are available in two formats indicated by their file extensions: .jmp and .csv. Files with the .jmp extension are proprietary to the JMP© statistics program from SAS Institute. These files contain the data and as well as information about variable coding, variable values, value ordering, and other information in column notes. In effect, the .jmp files contain the data and the code book. The .csv files are generally accessible for import into a wide variety of analytical software but contain no explanatory notes.

Finally, an annotated version of the on-line questionnaire is available as Appendix F of the original report from California (Kurani et al 2016) cited above. The on-line instrument is customized to each respondent as they complete it. More than simple skip patterns, as respondents answer questions content of subsequent questions is populated with information participants provide. Some of this requires calls to data external to the survey instrument; some of these data are proprietary and some are no longer available. Therefore, no "live" version of the on-line questionnaire from 2014 is maintained. The annotated version and the description of the survey provided in the linked report are provided to assist data users.

While household ownership and purchase of all light-duty passenger cars and trucks approach gender parity, to date zero emission vehicles (ZEVs) are being purchased by far more men than women. Prior analysis of data from California finds no reason based in the prospective interest in ZEVs of female and male respondents why this difference should persist. The present report extends the California analysis to 12 other US states with varying ZEV policy and market contexts.

Among many other contextual, socio-economic, demographic, and attitudinal measures, the survey solicited participants' prospective interest in acquiring an ZEV, that is, their interest in their next new car. Participants then indicated why they were motivated to select a ZEV or what motivated them to not select one. Factor analysis was used to reduce the dimensionality of participants' prior awareness, experience, knowledge, and assessments of ZEVs. Via nominal logistic regression modeling, differences in prospective interest in ZEVs between female and male respondents are examined. Given their prospective interest, the motivations of female and male respondents are compared.

Overall, no difference between female and male participants in prospective interest in a ZEV rises to the level of the observed differences in real markets. Further, the multivariate modeling indicates no statistically significant effect of a sex indicator on prospective interest in ZEVS almost anywhere in these states. Where there is a difference, female participants are estimated to be more likely to choose a ZEV than their male counterparts.

While participants from both sexes tend to give high scores to the same ZEV (de)motivations, differences in their rank orders repeat generalizations from other research. On average, female respondents score environmental motivations higher than do male respondents. On average, interest in "new technology" is more motivating to male than female participants. Conversely, on average female respondents who do not select a ZEV score "unfamiliar technology" more highly than their male counterparts.

Within the variation in policy and market contexts represented by the states in this study, no finding here explains why similar prospective interest among female and male participants in ZEVs from the beginning of 2015 has yet to be turned toward equal participation in ZEV markets. Explanations may lie in factors not modeled here.

Our dataset provides detailed and precise insights into the business, commercial, and industrial aspects of any given area in the USA (Including Point of Interest (POI) Data and Foot Traffic. The dataset is divided into 150x150 sqm areas (geohash 7) and has over 50 variables. - Use it for different applications: Our combined dataset, which includes POI and foot traffic data, can be employed for various purposes. Different data teams use it to guide retailers and FMCG brands in site selection, fuel marketing intelligence, analyze trade areas, and assess company risk. Our dataset has also proven to be useful for real estate investment.- Get reliable data: Our datasets have been processed, enriched, and tested so your data team can use them more quickly and accurately.- Ideal for trainning ML models. The high quality of our geographic information layers results from more than seven years of work dedicated to the deep understanding and modeling of geospatial Big Data. Among the features that distinguished this dataset is the use of anonymized and user-compliant mobile device GPS location, enriched with other alternative and public data.- Easy to use: Our dataset is user-friendly and can be easily integrated to your current models. Also, we can deliver your data in different formats, like .csv, according to your analysis requirements. - Get personalized guidance: In addition to providing reliable datasets, we advise your analysts on their correct implementation.Our data scientists can guide your internal team on the optimal algorithms and models to get the most out of the information we provide (without compromising the security of your internal data).Answer questions like: - What places does my target user visit in a particular area? Which are the best areas to place a new POS?- What is the average yearly income of users in a particular area?- What is the influx of visits that my competition receives?- What is the volume of traffic surrounding my current POS?This dataset is useful for getting insights from industries like:- Retail & FMCG- Banking, Finance, and Investment- Car Dealerships- Real Estate- Convenience Stores- Pharma and medical laboratories- Restaurant chains and franchises- Clothing chains and franchisesOur dataset includes more than 50 variables, such as:- Number of pedestrians seen in the area.- Number of vehicles seen in the area.- Average speed of movement of the vehicles seen in the area.- Point of Interest (POIs) (in number and type) seen in the area (supermarkets, pharmacies, recreational locations, restaurants, offices, hotels, parking lots, wholesalers, financial services, pet services, shopping malls, among others). - Average yearly income range (anonymized and aggregated) of the devices seen in the area.Notes to better understand this dataset:- POI confidence means the average confidence of POIs in the area. In this case, POIs are any kind of location, such as a restaurant, a hotel, or a library. - Category confidences, for example"food_drinks_tobacco_retail_confidence" indicates how confident we are in the existence of food/drink/tobacco retail locations in the area. - We added predictions for The Home Depot and Lowe's Home Improvement stores in the dataset sample. These predictions were the result of a machine-learning model that was trained with the data. Knowing where the current stores are, we can find the most similar areas for new stores to open.How efficient is a Geohash?Geohash is a faster, cost-effective geofencing option that reduces input data load and provides actionable information. Its benefits include faster querying, reduced cost, minimal configuration, and ease of use.Geohash ranges from 1 to 12 characters. The dataset can be split into variable-size geohashes, with the default being geohash7 (150m x 150m).

This deep learning model is used to detect cars in high resolution drone or aerial imagery. Car detection can be used for applications such as traffic management and analysis, parking lot utilization, urban planning, etc. It can also be used as a proxy for deriving economic indicators and estimating retail sales. High resolution aerial and drone imagery can be used for car detection due to its high spatio-temporal coverage.Using the modelFollow the guide to use the model. Before using this model, ensure that the supported deep learning libraries are installed. For more details, check Deep Learning Libraries Installer for ArcGIS.Fine-tuning the modelThis model can be fine-tuned using the Train Deep Learning Model tool. Follow the guide to fine-tune this model.InputHigh resolution RGB imagery (5 - 20 centimeter spatial resolution).OutputFeature class containing detected cars.Applicable geographiesThe model is expected to work well in the United States.Model architectureThis model uses the MaskRCNN model architecture implemented in ArcGIS API for Python.Accuracy metricsThis model has an average precision score of 0.81.Training dataThis model has been trained on an Esri proprietary car detection dataset.Sample resultsHere are a few results from the model. To view more, see this story.

This layer shows household size by number of vehicles available. This is shown by tract, county, and state centroids. This service is updated annually to contain the most currently released American Community Survey (ACS) 5-year data, and contains estimates and margins of error. There are also additional calculated attributes related to this topic, which can be mapped or used within analysis. This layer is symbolized to show the count and percentage of households with no vehicle available. To see the full list of attributes available in this service, go to the "Data" tab, and choose "Fields" at the top right. Current Vintage: 2019-2023ACS Table(s): B08201 Data downloaded from: Census Bureau's API for American Community Survey Date of API call: December 12, 2024National Figures: data.census.govThe United States Census Bureau's American Community Survey (ACS):About the SurveyGeography & ACSTechnical DocumentationNews & UpdatesThis ready-to-use layer can be used within ArcGIS Pro, ArcGIS Online, its configurable apps, dashboards, Story Maps, custom apps, and mobile apps. Data can also be exported for offline workflows. For more information about ACS layers, visit the FAQ. Please cite the Census and ACS when using this data.Data Note from the Census:Data are based on a sample and are subject to sampling variability. The degree of uncertainty for an estimate arising from sampling variability is represented through the use of a margin of error. The value shown here is the 90 percent margin of error. The margin of error can be interpreted as providing a 90 percent probability that the interval defined by the estimate minus the margin of error and the estimate plus the margin of error (the lower and upper confidence bounds) contains the true value. In addition to sampling variability, the ACS estimates are subject to nonsampling error (for a discussion of nonsampling variability, see Accuracy of the Data). The effect of nonsampling error is not represented in these tables.Data Processing Notes:This layer is updated automatically when the most current vintage of ACS data is released each year, usually in December. The layer always contains the latest available ACS 5-year estimates. It is updated annually within days of the Census Bureau's release schedule. Click here to learn more about ACS data releases.Boundaries come from the US Census TIGER geodatabases, specifically, the National Sub-State Geography Database (named tlgdb_(year)_a_us_substategeo.gdb). Boundaries are updated at the same time as the data updates (annually), and the boundary vintage appropriately matches the data vintage as specified by the Census. These are Census boundaries with water and/or coastlines erased for cartographic and mapping purposes. For census tracts, the water cutouts are derived from a subset of the 2020 Areal Hydrography boundaries offered by TIGER. Water bodies and rivers which are 50 million square meters or larger (mid to large sized water bodies) are erased from the tract level boundaries, as well as additional important features. For state and county boundaries, the water and coastlines are derived from the coastlines of the 2023 500k TIGER Cartographic Boundary Shapefiles. These are erased to more accurately portray the coastlines and Great Lakes. The original AWATER and ALAND fields are still available as attributes within the data table (units are square meters).The States layer contains 52 records - all US states, Washington D.C., and Puerto RicoCensus tracts with no population that occur in areas of water, such as oceans, are removed from this data service (Census Tracts beginning with 99).Percentages and derived counts, and associated margins of error, are calculated values (that can be identified by the "_calc_" stub in the field name), and abide by the specifications defined by the American Community Survey.Field alias names were created based on the Table Shells file available from the American Community Survey Summary File Documentation page.Negative values (e.g., -4444...) have been set to null, with the exception of -5555... which has been set to zero. These negative values exist in the raw API data to indicate the following situations:The margin of error column indicates that either no sample observations or too few sample observations were available to compute a standard error and thus the margin of error. A statistical test is not appropriate.Either no sample observations or too few sample observations were available to compute an estimate, or a ratio of medians cannot be calculated because one or both of the median estimates falls in the lowest interval or upper interval of an open-ended distribution.The median falls in the lowest interval of an open-ended distribution, or in the upper interval of an open-ended distribution. A statistical test is not appropriate.The estimate is controlled. A statistical test for sampling variability is not appropriate.The data for this geographic area cannot be displayed because the number of sample cases is too small.

United States Average Transaction Price: Manufacturer: Toyota Motor Sales USA, Inc. data was reported at 44,234.000 USD in Jan 2025. This records a decrease from the previous number of 44,981.000 USD for Dec 2024. United States Average Transaction Price: Manufacturer: Toyota Motor Sales USA, Inc. data is updated monthly, averaging 40,719.000 USD from Jan 2020 (Median) to Jan 2025, with 61 observations. The data reached an all-time high of 45,024.000 USD in Jul 2024 and a record low of 34,013.000 USD in Mar 2020. United States Average Transaction Price: Manufacturer: Toyota Motor Sales USA, Inc. data remains active status in CEIC and is reported by Cox Automotive. The data is categorized under Global Database’s United States – Table US.RA011: New Vehicle Average Transaction Price.

NHTSA's Corporate Average Fuel Economy (CAFE) program requires manufacturers of passenger cars and light trucks, produced for sale in the U.S., to meet CAFE standards, expressed in miles per gallon (mpg). The purpose of the CAFE program is to reduce the nation's energy consumption by increasing the fuel economy of cars and light trucks. The CAFE Public Information Center (PIC) is the authoritative source for Corporate Average Fuel Economy (CAFE) program data. This site allows fuel economy data to be viewed in report and/or graph format. The data can be sorted and filtered to produce custom reports which can also be downloaded as Excel or pdf files. NHTSA periodically updates the CAFE data in the PIC and, therefore, each report and graph is date stamped to indicate the last time NHTSA made updates.

How many people are staying at home? How far are people traveling when they don’t stay home? Which states and counties have more people taking trips? The Bureau of Transportation Statistics (BTS) now provides answers to those questions through our mobility statistics program.

The "Trips by Distance" data and number of people staying home and not staying home are estimated for the Bureau of Transportation Statistics by the Maryland Transportation Institute and Center for Advanced Transportation Technology Laboratory at the University of Maryland. The travel statistics are produced from an anonymized national panel of mobile device data from multiple sources. All data sources used in the creation of the metrics contain no personal information. Data analysis is conducted at the aggregate national, state, and county levels. A weighting procedure expands the sample of millions of mobile devices, so the results are representative of the entire population in a nation, state, or county. To assure confidentiality and support data quality, no data are reported for a county if it has fewer than 50 devices in the sample on any given day.

Trips are defined as movements that include a stay of longer than 10 minutes at an anonymized location away from home. Home locations are imputed on a weekly basis. A movement with multiple stays of longer than 10 minutes before returning home is counted as multiple trips. Trips capture travel by all modes of transportation. including driving, rail, transit, and air.

The daily travel estimates are from a mobile device data panel from merged multiple data sources that address the geographic and temporal sample variation issues often observed in a single data source. The merged data panel only includes mobile devices whose anonymized location data meet a set of data quality standards, which further ensures the overall data quality and consistency. The data quality standards consider both temporal frequency and spatial accuracy of anonymized location point observations, temporal coverage and representativeness at the device level, spatial representativeness at the sample and county level, etc. A multi-level weighting method that employs both device and trip-level weights expands the sample to the underlying population at the county and state levels, before travel statistics are computed.

These data are experimental and may not meet all of our quality standards. Experimental data products are created using new data sources or methodologies that benefit data users in the absence of other relevant products. We are seeking feedback from data users and stakeholders on the quality and usefulness of these new products. Experimental data products that meet our quality standards and demonstrate sufficient user demand may enter regular production if resources permit.

These data are made available under a public domain license. Data should be attributed to the "Maryland Transportation Institute and Center for Advanced Transportation Technology Laboratory at the University of Maryland and the United States Bureau of Transportation Statistics."

Daily data for a given week will be uploaded to the BTS website within 9-10 days of the end of the week in question (e.g., data for Sunday September 17-Saturday September 23 would be updated on Tuesday, October 3). All BTS visualizations and tables that rely on these data will update at approximately 10am ET on days when new data are received, processed, and uploaded.

The methodology used to develop these data can be found at: https://rosap.ntl.bts.gov/view/dot/67520.

Simulated hourly electric vehicle charging profiles for light-duty household passenger vehicles in the contiguous United States, 2018-2050. Profiles are differentiated by scenario, county, household and vehicle types, and charging type. Data was produced in 2022 using the Transportation Energy & Mobility Pathway Options (TEMPO) model and published in demand-side grid (dsgrid) toolkit format.

Data are available for three adoption scenarios: "AEO Reference Case", which is aligned with the U.S. EIA Annual Energy Outlook 2018 (linked below), "EFS High Electrification", which is aligned with the High Electrification scenario of the Electrification Futures Study (linked below), and "All EV Sales by 2035", which assumes that average passenger light-duty EV sales reach 50% in 2030 and 100% in 2035.

The charging shapes are derived from two key assumptions of which data users should be aware: "ubiquitous charger access", meaning that drivers of vehicles are assumed to have access to a charger whenever a trip is not in progress, and "immediate charging", meaning that immediately after trip completion, vehicles are plugged in and charge until they are either fully recharged or taken on another trip.

These assumptions result in a bounding case in which vehicles' state of charge is maximized at all times. This bounding case would minimize range anxiety, but is unrealistic from the point of view of both electric vehicle service equipment (EVSE) (i.e., charger) access, and plug-in behavior as it can result in dozens of charging sessions per week for battery electric vehicles (BEVs) that in reality are often only plugged in a few times per week.

Dataset Statistics: The statistics of our dataset are summarized and compared with the largest existing dataset (DR(eye)VE) [1] in Table 1. Our dataset was collected using videos selected from a publicly available, large-scale, crowd-sourced driving video dataset, BDD100k [30, 31]. BDD100K contains human-demonstrated dashboard videos and time-stamped sensor measurements collected during urban driving in various weather and lighting conditions. To efficiently collect attention data for critical driving situations, we specifically selected video clips that both included braking events and took place in busy areas (see supplementary materials for technical details). We then trimmed videos to include 6.5 seconds prior to and 3.5 seconds after each braking event. It turned out that other driving actions, e.g., turning, lane switching and accelerating, were also included. 1,232 videos (=3.5 hours) in total were collected following these procedures. Some example images from our dataset are shown in Fig. 6. Our selected videos contain a large number of different road users. We detected the objects in our videos using YOLO [22].On average, each video frame contained 4.4 cars and 0.3 pedestrians, multiple times more than the DR(eye)VE dataset (Table 1). Data Collection Procedure: For our eye-tracking experiment, we recruited 45 participants who each had more than one year of driving experience. The participants watched the selected driving videos in the lab while performing a driving instructor task: participants were asked to imagine that they were driving instructors sitting in the copilot seat and needed to press the space key whenever they felt it necessary to correct or warn the student driver of potential dangers. Their eye movements during the task were recorded at 1000 Hz with an EyeLink 1000 desktop-mounted infrared eye tracker, used in conjunction with the Eyelink Toolbox scripts [7] for MATLAB. Each participant completed the task for 200 driving videos. Each driving video was viewed by at least 4 participants. The gaze patterns made by these independent participants were aggregated and smoothed to make an attention map for each frame of the stimulus video (see Fig. 6 and supplementary materials for technical details). Psychological studies [19, 11] have shown that when humans look through multiple visual cues that simultaneously demand attention, the order in which humans look at those cues is highly subjective. Therefore, by aggregating gazes of independent observers, we could record multiple important visual cues in one frame. In addition, it has been shown that human drivers look at buildings, trees, flowerbeds, and other unimportant objects non-negligibly frequently [1]. Presumably, these eye movements should be regarded as noise for driving-related machine learning purposes. By averaging the eye movements of independent observers, we were able to effectively wash out those sources of noise (see Fig. 2B). Comparison with In-Car Attention Data: We collected in-lab driver attention data using videos from the DR(eye)VE dataset. This allowed us to compare in-lab and in-car attention maps of each video. The DR(eye)VE videos we used were 200 randomly selected 10-second video clips, half of them containing braking events and half without braking events. We tested how well in-car and in-lab attention maps highlighted driving-relevant objects. We used YOLO [22] to detect the objects in the videos of our dataset. We identified three object categories that are important for driving and that had sufficient instances in the videos (car, pedestrian and cyclist). We calculated the proportion of attended objects out of total detected instances for each category for both in-lab and in-car attention maps (see supplementary materials for technical details). The results showed that in-car attention maps highlighted significantly less driving-relevant objects than in-lab attention maps (see Fig. 2A). The difference in the number of attended objects between the in-car and in-lab attention maps can be due to the fact that eye movements collected from a single driver do not completely indicate all the objects that demand attention in the particular driving situation. One individual’s eye movements are only an approximation of their attention [23], and humans can also track objects with covert attention without looking at them [6]. The difference in the number of attended objects may also reflect the difference between first-person driver attention and third-person driver attention. It may be that the human observers in our in-lab eye-tracking experiment also looked at objects that were not relevant for driving. We ran a human evaluation experiment to address this concern. Human Evaluation: To verify that our in-lab driver attention maps highlight regions that should indeed demand drivers’ attention, we conducted an online study to let humans compare in-lab and in-car driver attention maps. In each trial of the online study, participants watched one driving video clip three times: the first time with no edit, and then two more times in random order with overlaid in-lab and in-car attention maps, respectively. The participant was then asked to choose which heatmap-coded video was more similar to where a good driver would look. In total, we collected 736 trials from 32 online participants. We found that our in-lab attention maps were more often preferred by the participants than the in-car attention maps (71% versus 29% of all trials, statistically significant as p = 1×10−29, see Table 2). Although this result cannot suggest that in-lab driver attention maps are superior to in-car attention maps in general, it does show that the driver attention maps collected with our protocol represent where a good driver should look from a third-person perspective. In addition, we will show in the Experiments section that in-lab attention data collected using our protocol can be used to train a model to effectively predict actual, in-car driver attention. This result proves that our dataset can also serve as a substitute for in-car driver attention data, especially in crucial situations where in-car data collection is not practical. To summarize, compared with driver attention data collected in-car, our dataset has three clear advantages: multi-focus, little driving-irrelevant noise, and efficiently tailored to crucial driving situations.