PaperlayerView exact revision on arXiv

Source-linked AI summary

DADA-2000: Can Driving Accident be Predicted by Driver Attention? Analyzed by A Benchmark

Jianwu Fang, Dingxin Yan, Jiahuan Qiao, Jianru Xue, He Wang, Sen Li

arXiv:1904.12634v1cs.CVcs.AI

TL;DR

Prior attention datasets do not closely connect attention in critical situations to prediction of actual accidents, an important task for safe driving. This paper constructs DADA-2000, a diverse benchmark pairing driver-attention data with accident annotations. Its analysis reports that attention can play a positive role in crash-object localization and gives a positive answer to the accident-prediction question.

  • Problem

    BDD-A focuses on critical situations that do not become actual accidents, leaving the dynamic attention process from critical situations to actual accidents unexplored.

    Source: §I · p. 2

  • Method

    The authors construct DADA-2000, pairing driver-attention data with accident categories, temporal windows, and crash-object locations across 2,000 clips.

    Source: §I · p. 2

  • Results

    About 80% of abnormal frames have CLE below 300 pixels, and the analysis indicates driver attention can play a positive role in crash-object localization.

    Source: 1) Overall Analysis · p. 6

  • Takeaways & Limitations

    The authors report a quantitative, positive answer to whether driving accidents can be predicted by driver attention.

    Source: §V · p. 7

Abstract

from arXiv · show

Driver attention prediction is currently becoming the focus in safe driving research community, such as the DR(eye)VE project and newly emerged Berkeley DeepDrive Attention (BDD-A) database in critical situations. In safe driving, an essential task is to predict the incoming accidents as early as possible. BDD-A was aware of this problem and collected the driver attention in laboratory because of the rarity of such scenes. Nevertheless, BDD-A focuses the critical situations which do not encounter actual accidents, and just faces the driver attention prediction task, without a close step for accident prediction. In contrast to this, we explore the view of drivers' eyes for capturing multiple kinds of accidents, and construct a more diverse and larger video benchmark than ever before with the driver attention and the driving accident annotation simultaneously (named as DADA-2000), which has 2000 video clips owning about 658,476 frames on 54 kinds of accidents. These clips are crowd-sourced and captured in various occasions (highway, urban, rural, and tunnel), weather (sunny, rainy and snowy) and light conditions (daytime and nighttime). For the driver attention representation, we collect the maps of fixations, saccade scan path and focusing time. The accidents are annotated by their categories, the accident window in clips and spatial locations of the crash-objects. Based on the analysis, we obtain a quantitative and positive answer for the question in this paper.

I. INTRODUCTION

Driver attention is studied as a route to safer driving, but existing datasets leave gaps between attention in critical situations and prediction of actual accidents. DADA-2000 combines driver-attention and accident annotations across diverse accident clips, and the paper reports a positive answer to whether attention can predict accidents.

I. INTRODUCTION

  • Driver attention matters for safe-driving research because eye movements can respond immediately to critical driving situations.The introduction links attention to visual scene search and interaction with surroundings.Source: §I · p. 1
  • BDD-A covers critical situations without actual accidents, leaving the dynamic attention process from critical situations to real accidents unexplored.The authors also note that BDD-A averages away attention to multiple observers’ non-salient objects and does not classify braking events into fine-grained categories.Source: §I · pp. 1–2
  • DADA-2000 combines driver-attention and accident annotations in a larger, more diverse video benchmark spanning 54 accident types and 2,000 clips.The 658,476 selected frames come from crowd-sourced videos across varied weather, lighting, and scene occasions; attention is collected in a lab protocol and crash-object locations and accident intervals are annotated.Source: §I · p. 2
  • The paper reports a quantitative, positive answer to whether driving accidents can be predicted from driver attention.The paper also frames the benchmark as a step toward exploring attention flow from normal situations to accidents.Source: §I · p. 2

II. RELATED WORKS

Prior driver-attention datasets cover ordinary driving or critical situations, while accident-prediction work generally focuses on temporal prediction or detection. The paper positions DADA-2000 as a broader attention benchmark that also captures actual accidents.

II. RELATED WORKS

  • Prior driver-attention datasets leave actual accident situations or the dynamic attention process leading to accidents unaddressed.DR(eye)VE focuses on unobstructed scenes, while BDD-A does not capture the process from critical situations to actual accidents or categorize braking events into subclasses.Source: §II · p. 2
  • DADA-2000 is presented as a larger driver-attention benchmark covering more diverse driving situations than the prior attention datasets discussed.Its annotations address attention alongside accident categories, temporal windows, and crash-object locations.Source: §II · p. 2
  • Earlier accident-prediction studies generally target temporal interval prediction or detection, without spatial crash-object localization.The reviewed methods also rely on expensive optical flow and may lack time to respond to sudden accidents.Source: §II · p. 2

III. DADA-2000

DADA-2000 was assembled from crowd-sourced accident videos and annotated with accident categories, crash-object locations, and temporal windows. Its temporal partition separates the accident window from the periods before and after it.

III. DADA-2000

  • The dataset retains 658,476 frames from 2,000 videos after about 3 million downloaded frames were screened for usable content.The clips have 1456 × 660 resolution and total about 6.1 hours at 30 fps.Source: §III · p. 3
  • Videos are presented without trimming and classified into 54 accident categories based on participants, grouped as ego-car-involved or ego-car-uninvolved.The paper also provides 62 categories in its category illustration; ego-car hitting car is the largest proportion.Source: §III · p. 3
  • Each clip is annotated with crash-object locations, an accident window, and frame-level attention maps.The accident window starts when half of the future crash-object appears in view and ends when the scene returns to normal movement.Source: §III · pp. 3–4
  • The accident window is divided into start, middle, and end sections, with the earlier two intended to examine whether attention can capture a crash-object early.The start and end of the window are marked by the crash-object entering view and the scene returning to normal movement, respectively.Source: §III · p. 4

1) Temporal Statistics:

Temporal statistics show that accident windows occupy a minority of each clip. Crash-object locations vary by accident category, while the paper describes DADA-2000 as a more diverse and complex accident-analysis dataset than prior datasets.

1) Temporal Statistics:

  • 20.3% of each video is the accident window on average, compared with about 5 seconds before it and 3.5 seconds after it.The paper concludes that abnormal-driving frames are fewer than normal-driving frames.Source: 1) Temporal Statistics · p. 4

2) Spatial Statistics:

  • Crash-object locations converge toward the middle of the field of vision for most accident categories, but crossing categories tend toward its sides.The spatial patterns are summarized in occurrence maps of crash-object locations.Source: 2) Spatial Statistics · p. 4
  • Hitting behaviors converge toward the middle of the field of vision, with stronger convergence as target velocity increases.The paper notes that cars show greater convergence toward the sides than motorbikes and pedestrians in crossing cases.Source: 2) Spatial Statistics · p. 4
  • DADA-2000 is described as having more diverse scenarios and greater complexity than existing driving-accident datasets.The authors present it as a larger platform for accident detection or prediction research.Source: 2) Spatial Statistics · p. 4

B. Attention Collection

Attention data were collected in a laboratory protocol with repeated viewing conditions and at least five observers per video. The benchmark preserves attention in two forms, including unaggregated group fixations and temporally aggregated individual maps.

1) Protocols:

  • Twenty experienced volunteers recorded eye movements in a dimmed laboratory, and each clip was viewed by at least five observers.Each volunteer watched sequences of 40 clips, grouped by category where possible to avoid chaotic attention.Source: 1) Protocols · p. 4
  • Volunteers watched each video twice: once without being told to find the crash-object and once with that instruction.These conditions support analysis of unconscious and conscious crash attention, respectively.Source: 1) Protocols · p. 4

2) Statistics:

  • Type1 maps retain five observers’ fixations together without temporal aggregation, while Type2 maps preserve each observer individually with a one-second temporal window.The authors report that unaggregated fixations can reveal the common focusing effect, while the aggregated maps retain dynamic attention information.Source: 2) Statistics · pp. 4–5
  • Figure 6 presents spatial occurrence maps for typical accident categories and for DADA-2020.Bracketed category numbers refer to the category index in Figure 2.Source: Figure 6 · p. 5
  • The paper presents DADA-2000 as more diverse than existing driver-attention datasets and includes an attribute comparison of those datasets.The comparison is identified in Table III.Source: 2) Statistics · p. 5, Table III · p. 5

IV. CAN DRIVING ACCIDENT BE PREDICTED BY DRIVER ATTENTION?

The paper asks whether driver attention can predict driving accidents. It motivates the question by the insufficiency of prior accident-prediction performance for safe driving and the lack of a large-scale attention benchmark for accident analysis.

IV. CAN DRIVING ACCIDENT BE PREDICTED BY DRIVER ATTENTION?

  • The central question is whether driver attention can predict driving accidents.The authors frame this as a search for a new solution to accident prediction or detection.Source: §IV · p. 5
  • Prior accident-prediction or detection paradigms often use complex models, yet their performance remains insufficient for safe driving [18].Source: §IV · p. 5
  • The authors identify the absence of a promising large-scale attention benchmark for driving-accident analysis, where sudden motion changes are frequent.Source: §IV · p. 5

A. Attended Object Categories

The analysis compares attended object categories in normal and abnormal driving, with and without a prior. In abnormal situations, attention shifts toward vulnerable road users, while the prior has little influence on attended objects.

A. Attended Object Categories

  • The analysis compares top-15 object-category proportions in normal driving before and after the accident window with those during the window.Semantic maps for each frame were generated using Deeplab-V3 pretrained on ADE20K.Source: Figure 7 · p. 5, A. Attended Object Categories · pp. 5–6
  • In abnormal driving, person shifts from tenth to fifth in attended-object rank and cars move to first, indicating greater attention to vulnerable road users.The analysis finds little influence of the prior on which objects drivers attend.Source: A. Attended Object Categories · p. 6

B. Localization Error Analysis

The section evaluates crash-object localization using center location error (CLE), the distance between the crash-object center and the peak saliency point. Across accident-window frames, most CLE values are below 300 pixels, a threshold the authors contextualize against crash-object size and dataset composition.

B. Localization Error Analysis

1) Overall Analysis:

  • About 80% of abnormal frames (131,679 total) have CLE below 300 pixels, which the authors regard as acceptable given crash-object size and dataset composition.Ego-car-hitting-car accidents are the largest category, and ego-car-involved categories comprise almost half of DADA-2000.Source: 1) Overall Analysis · p. 6, Figure 9 · p. 6
  • The example with the largest crash-object shows a crash car occupying nearly half of the image plane.Source: Figure 9 · p. 6

2) Analysis for Different Sections in AW:

The section evaluates crash-object localization across accident-window sections using frame ratio and clip-level success rate. Attention localizes objects in all sections, with the authors reporting strong localization at a 300-pixel CLE threshold and early capture in the start section.

2) Analysis for Different Sections in AW:

V. CONCLUSIONS

The conclusion presents DADA-2000 as a benchmark for testing whether driver attention can help predict driving accidents. Quantitative analysis gives a positive answer to that question.

V. CONCLUSIONS

  • DADA-2000 contains 2,000 video clips with over 650,000 frames, making it larger than the related datasets described by the authors.Source: §V · p. 7
  • The benchmark annotates temporal accident windows and spatial crash-object locations across its videos, alongside driver-attention data for each frame.Source: §V · p. 7
  • Quantitative analysis yields a positive answer to whether driver attention can help predict driving accidents.Source: §V · p. 7

Related papers

Loading 1904.12634v1…