SPEED+: Leading-Vehicle Velocity Recognition Dataset

• SPEED+ is an open dataset designed for robust leading-vehicle velocity recognition in dynamic following scenarios under adverse weather and low-light conditions.
• It leverages time-series stereo camera images and supplementary sensor labels, including bounding boxes and steering angle, to support precise vehicle motion estimation.
• The benchmark's evaluation metric emphasizes prompt tracking and timely prediction, ensuring safety-critical performance for advanced driver-assistance systems.

Searching arXiv for papers mentioning SPEED+ to verify disambiguation across domains. SPEED+ is an open dataset and benchmark for robust leading-vehicle velocity recognition from camera-based in-vehicle sensing, designed for Advanced Driver-Assistance Systems and Autonomous Driving under challenging real-world conditions such as nighttime and rain (Ogawa et al., 2022). It targets the time-series task of estimating the absolute velocity of the vehicle being followed from sequences of stereo camera images and related vehicle information. The dataset is application-specific: it focuses on following scenarios, uses images from a commercial stereo camera system, and provides labels not only for target velocity but also for supporting quantities such as the bounding-box position of the leading vehicle in the image, distance to the leading vehicle, ego-vehicle velocity, steering angle, and stereo disparity information.

1. Definition and task formulation

SPEED+ is built around the task of recognizing the velocity of the leading vehicle in front of the ego vehicle. The task is important for functions such as automatic emergency braking and adaptive cruise control, because it requires accurate estimation of the motion of the vehicle ahead in order to avoid rear-end collisions and maintain safe following distance (Ogawa et al., 2022). The paper frames the problem as a time-series recognition problem in which a model receives a sequence of stereo camera images and related information and must estimate the absolute velocity of the vehicle being followed.

The task is narrower than generic driving-scene understanding. It involves multiple subproblems—detecting the leading vehicle, tracking it across frames, estimating its position in 3D, and inferring its velocity—but the benchmark is centered specifically on leading-vehicle velocity recognition. The inclusion of supporting labels makes the dataset usable not only for end-to-end velocity prediction but also for studying intermediate perception components such as detection and tracking.

A common misconception is that SPEED+ is a general-purpose autonomous-driving dataset. It is not. The paper explicitly states that the dataset focuses on following scenarios in which the leading vehicle is within 120 m, and it is specialized for a practical closed perception problem relevant to AEB and ACC rather than for a broad mixture of detection, segmentation, tracking, and mapping tasks (Ogawa et al., 2022).

2. Motivation and position relative to earlier datasets

The dataset was created because existing public resources were not considered ideal for this specific problem. Many public datasets were described as being captured in ordinary driving conditions and therefore lacking enough adverse weather and night scenes. DAWN was cited as focusing on adverse weather, but using web-collected images that are not well suited to development for fixed on-board camera systems. KITTI was described as a major benchmark containing a mixture of tasks such as detection, tracking, and 3D recognition, but not as a dataset directly tailored to the practical development of functions like AEB and ACC (Ogawa et al., 2022).

SPEED+ was therefore designed to be closer to practical use. The paper emphasizes three points in this positioning: it uses images from a commercial stereo camera system, it includes many difficult scenes, and it provides a benchmark focused on an application-critical vehicle motion task. This suggests that the dataset is intended less as a broad scene-understanding corpus than as a controlled benchmark for a specific perception capability that sits close to production deployment.

The paper also distinguishes SPEED+ from generic vision datasets such as ImageNet, MS COCO, Cityscapes, Mapillary, and BDD100K. Those datasets are useful for detection and segmentation, but the authors argue that they mostly contain normal driving scenes and do not provide enough adverse camera conditions. In that sense, SPEED+ is application-specific and condition-enriched rather than task-general.

3. Sensor configuration and dataset composition

SPEED+ is camera-based and specifically uses a stereo camera from Subaru’s EyeSight system (Ogawa et al., 2022). The paper gives the image specifications as follows.

Component	Specification	Notes
Right camera image	1000 × 420 pixels	Stereo input
Left camera image	1128 × 420 pixels	Stereo input
Frame rate	10 fps	Time-series benchmark
Scenes	1000 scenes	Following scenes
Frames per scene	100 to 200 frames	Temporal sequences

All scenes are those in which the ego vehicle is following a leading vehicle, and scenes were selected so that the leading vehicle was within 120 m. The video data were compressed using mp4 to reduce size, and the authors report that this compression did not significantly affect general detection performance. For public release, license plates and pedestrian faces were blurred, and the authors verified that this did not significantly affect recognition performance; they also checked that no readable house nameplates remained (Ogawa et al., 2022).

The dataset’s scene distribution includes time of day, weather, driving place, and leading vehicle type. Specifically, the paper describes categories of daytime / night / other, fine/cloudy / rain / other, highway / ordinary road, and car / bus-truck. The authors state that they considered the distribution of typical driver scenes but intentionally included more night and rain cases because these are especially challenging for vision systems. Difficult scenes with water droplets on the windshield and wiper occlusion, especially at night or in rain, are highlighted as important real-world edge cases.

This composition matters because robustness under adverse image conditions is one of the principal stated goals of SPEED+. A plausible implication is that the benchmark is meant to stress-test visual temporal models under the kinds of photometric and occlusion effects that are often underrepresented in conventional driving datasets.

4. Labels and ground-truth construction

The dataset contains several labels per frame: the leading vehicle position in the image, distance to the leading vehicle, leading vehicle velocity, ego-vehicle velocity, steering angle, and a parameter to convert disparity to distance (Ogawa et al., 2022). The leading vehicle position is provided as a rectangular bounding box in the right camera image. In the training data, each frame has this position label, whereas in the test data only the first frame provides the leading vehicle position so as to specify which object should be tracked or predicted.

Distance to the leading vehicle is measured using IBEO laser radar. Leading vehicle velocity is derived from distance measurements. The paper states that velocity is calculated by differentiating distance and averaging over 7 frames before and after the current frame. Ego-vehicle velocity and steering angle are also included as vehicle information, and the disparity-related parameter supports stereo-based distance estimation.

The paper explains the ground-truth velocity definition in two stages. First, the distance to the leading vehicle is differentiated over time to obtain the relative velocity. Then the absolute velocity is computed by adding the ego vehicle’s velocity. An important caveat is explicitly noted: because the velocity label is generated using future frames through averaging frames before and after the target frame, the task partly requires future prediction in the benchmark setting (Ogawa et al., 2022). This means the target is not simply an instantaneous quantity available from past frames alone.

A common misunderstanding is that the label corresponds to a direct sensor reading of instantaneous lead-vehicle speed. The paper instead describes a derived and temporally smoothed target. That detail is not incidental; it materially affects how temporal context should be interpreted in model design and in benchmark evaluation.

5. Benchmark protocol and evaluation metric

SPEED+ was explicitly designed to support benchmarking, and the public evaluation was hosted on SIGNATE (Ogawa et al., 2022). The competition lasted 2.5 months, was open to individuals, organizations, and nationalities, and imposed constraints of memory no greater than 32 GB, GPU memory no greater than 10 GB, and model file size no greater than 2 GB. The reported participation figures were 126 teams and 2077 submissions. The top three scores were 0.0846, 0.0915, and 0.0917, and the paper states that for practical use an evaluation value under 0.1 was required.

The evaluation is not a plain regression metric. For scene $j$, with $n_j$ frames, predicted value $p_i$, and target value $t_i$ at frame $i$, the paper defines

$$e_j = \frac{1}{n_j - 19}\sum_{i=20}^{n_j}\min\left(\frac{|p_i - t_i|}{0.07\,t_i + 1},\,1\right).$$

The first 19 frames are excluded from evaluation because the task is time-series based and early frames are treated as warm-up with insufficient temporal context. The normalization term $0.07\,t_i + 1$ prevents high-speed scenes from dominating the metric, and the $\min(\cdot,1)$ operation caps the per-frame penalty.

The final score is a weighted average over scenes:

$\text{error} = \frac{\sum_j w_j e_j}{\sum_j w_j},$

with $w_j = 3$ if the distance to the leading vehicle is decreasing and $n_j$0 otherwise (Ogawa et al., 2022). The higher weight on decreasing-distance scenes reflects their safety-critical character.

The paper further interprets the metric in terms of two practical qualities. First, it is sensitive to following capability and recognition delay: simulated prediction delays of $n_j$1 frames degrade the metric, so the benchmark captures whether a model can follow the leading vehicle promptly. Second, it is less sensitive to stability than to delay, though stability still matters. This indicates that the metric is designed to emphasize timely tracking behavior, which the authors consider appropriate for ADAS applications.

6. Practical significance, uses, and limitations

The paper explicitly connects SPEED+ to automatic emergency braking, adaptive cruise control, broader ADAS, and autonomous driving (Ogawa et al., 2022). It is intended to help improve recognition robustness in difficult conditions, leading-vehicle motion estimation, and downstream safety functions. Because it includes bounding boxes, distance, ego-vehicle state, steering angle, and disparity-related information, it is also useful for developing subcomponents such as vehicle detection, tracking, and stereo-based 3D understanding.

Its practical orientation is reinforced by several design choices. The dataset is captured by a real commercial stereo camera system, enriched with night and rain scenes, and structured as a benchmark around a single application-critical perception task. The public competition further turns it into a comparative framework rather than merely a static collection of annotated sequences.

At the same time, the paper makes several limitations clear. The dataset is specialized to following scenarios with a leading vehicle within 120 m, so it is not fully general. The benchmark depends on stereo camera input, so methods relying on other modalities are not directly addressed. The ground-truth velocity is derived from sensor measurements and temporal smoothing, and some scenes in which laser radar had difficulty were excluded. The evaluation metric emphasizes timeliness and following capability more than pure smoothness (Ogawa et al., 2022).

A further misconception is that SPEED+ should be read primarily as a detection benchmark. The paper’s design indicates otherwise: detection and tracking are relevant, but they are intermediate components inside a benchmark whose principal target is robust leading-vehicle velocity recognition under production-like sensing and adverse visual conditions.

7. Broader interpretation

Within the ADAS and autonomous-driving literature, SPEED+ occupies a specific niche: it is narrower than large-scale generic driving datasets, but more tightly aligned with a practical perception function required by vehicle-following systems (Ogawa et al., 2022). Its emphasis on difficult scenes—nighttime, rain, water droplets on the windshield, and wiper occlusion—makes it a benchmark for robustness rather than only for nominal-condition accuracy.

The dataset also formalizes an unusual target-construction choice. Because the leading-vehicle velocity label is derived from differentiated distance and averaged over 7 frames before and after the current frame, the benchmark partly embeds future information into the target. This suggests that models are being evaluated not merely on instantaneous geometric inference, but on their ability to produce temporally aligned and behaviorally prompt estimates under a smoothed label definition.

Taken together, these properties make SPEED+ a specialized benchmark for a safety-relevant time-series recognition problem: estimating the absolute velocity of a leading vehicle from stereo in-vehicle camera data under difficult real-world conditions, with evaluation criteria that prioritize prompt following behavior in collision-relevant scenarios.