Visual crowding—the deleterious influence of nearby objects on object recognition—is considered to be a major bottleneck for object recognition in cluttered environments. Although crowding has been studied for decades with static and artificial stimuli, it is still unclear how crowding operates when viewing natural dynamic scenes in real-life situations. For example, driving is a frequent and potentially fatal real-life situation where crowding may play a critical role. In order to investigate the role of crowding in this kind of situation, we presented observers with naturalistic driving videos and recorded their eye movements while they performed a simulated driving task. We found that the saccade localization on pedestrians was impacted by visual clutter, in a manner consistent with the diagnostic criteria of crowding (Bouma's rule of thumb, flanker similarity tuning, and the radial-tangential anisotropy). In order to further confirm that altered saccadic localization is a behavioral consequence of crowding, we also showed that crowding occurs in the recognition of cluttered pedestrians in a more conventional crowding paradigm. We asked participants to discriminate the gender of pedestrians in static video frames and found that the altered saccadic localization correlated with the degree of crowding of the saccade targets. Taken together, our results provide strong evidence that crowding impacts both recognition and goal-directed actions in natural driving situations.

^{2}, and motion threshold = 0°).

*p*< 0.001). Note that the 95% percentile of the duration of the direct saccades was just 64 ms, which suggests that a delay of 12 ms is a significant amount.

*p*value was 0.96. Therefore, we ruled out target retinal size as a confounder and did not include it in the following analyses of Experiment 1.

*p*is the PAS, and α and β are fitted parameters. The α indicates the fitted PAS at an eccentricity of 2.5° and β quantifies how fast the PAS increases with increasing eccentricity. We compared the parameters fitted for flanked targets and unflanked targets, and the results confirmed our hypothesis (α

_{f}− α

_{u}= − 0.08, permutation test

*p*= 0.56; β

_{f}− β

_{u}= 0.088, permutation test

*p*= 0.006; Figure 3).

*p*= 0.36) and a significantly higher mean PAS for flanked targets than for unflanked targets for target eccentricity larger than 5° (permutation test

*p*= 0.01).

*p*= 0.02, see Figure 4). Within 5° eccentricity, there was no significant difference between car flankers and pedestrian flankers (permutation test

*p*= 0.37, see Figure 4). These results suggest that target-similar flankers were more effective at crowding the target, consistent with the well-known similarity tuning of crowding (Andriessen & Bouma, 1976; Chung, Levi, & Legge, 2001; Kooi et al., 1994; Levi, 2008).

_{<0.5}= − 0.33), and a relatively flat slope over the ratio range above 0.5 (β

_{≥0.5}= 0.03). Hence, the difference between the two slopes was consistent with Bouma's rule-of-thumb. However, target eccentricity could still be a confounder because larger target eccentricity leads to higher PAS (as seen in Figure 3) and the saccades of different spacing-to-eccentricity ratios have different mean target eccentricities. To determine the significance level of the difference in slopes (β

_{<0.5}− β

_{≥0.5}) after accounting for target eccentricity, we added target eccentricity as a regressor into the clipped line fitting model and conducted a permutation test where we shuffled the spacing-to-eccentricity ratio values of the saccades. Results showed a significant slope change (β

_{<0.5}− β

_{≥0.5}= − 0.36, permutation-test

*p*< 0.001; see Figure 5B), thus further confirming consistency with Bouma's rule.

*X*is a dummy variable that is equal to 1 when the flanker alignment of the saccade is radial and 0 when the flanker alignment is tangential,

_{radial}*X*is a dummy variable that is equal to 1 when the saccade is horizontal and 0 otherwise, and

_{horizontal}*LPSD*is logarithmic pictorial size difference. β quantifies the independent influence of radial versus tangential flanker alignment on PAS after accounting for the influence of target eccentricity, saccade direction, and target-flanker depth difference. The fitting showed that β = 0.93, permutation test

*p*= 0.003 (Figure 6B). The effect was significant even after accounting for target eccentricity.

*p*= 0.16). Overall, the result suggested that the influence of flanking pedestrians on saccade landing accuracy is consistent with the radial-tangential anisotropy.

^{○}and the standard deviation was 2.6

^{○}.

*accur*is the gender discrimination accuracy,

*eccen*is the eccentricity of the target, and α and β are fitted parameters. α indicates the fitted accuracy at an eccentricity of 2.5° and β quantifies how fast the accuracy decreases with increasing eccentricity. We compared the parameters fitted for flanked targets and unflanked targets, and the results followed our expectation (α

_{f}− α

_{u}= 0.016, permutation test

*p*= 0.92; β

_{f}− β

_{u}= − 0.22, permutation test

*p*< 0.001; see Figure 7).

*size*is the retinal size of the target pedestrian. This model captured how the accuracy varied based on target eccentricity and target retinal size, regardless of whether the target was flanked or not. For each trial, we then simulated the outcome of the trial (i.e. correct or wrong) based on a binomial distribution with the probability of correctness equal to the accuracy predicted by model 4. Simulated data showed how data would distribute under the null hypothesis. We then fit model 3 to the simulated data separately for flanked targets and unflanked targets to obtain the baseline accuracy-eccentricity curves under the null hypothesis. Baseline curves are plotted as dashed curves in Figure 7. To calculate the significance level of the difference of how fast accuracy decreased with eccentricity between flanked and unflanked targets after discounting the effect of target retinal size, we fit model 4 separately to the flanked data and unflanked data. This time β

_{f}− β

_{u}= − 0.24. We ran a permutation test where we shuffled the flanked/unflanked labels of all the trials to get a null distribution of β

_{f}− β

_{u}. The permutation test showed that

*p*< 0.001. Hence, we confirmed that even after accounting for the effect of target retinal size, gender discrimination accuracy dropped with increasing eccentricity significantly faster for flanked targets than for unflanked targets. Similarly, we confirmed that after discounting the effect of target retinal size, there was no significant difference in the accuracy at 2.5° eccentricity between flanked and unflanked targets (α

_{f}− α

_{u}= 0.23, permutation test

*p*= 0.49).

*accur*−

_{f}*accur*= − 0.05%), and a lower mean accuracy for flanked targets than for unflanked targets for target eccentricities beyond 5° (

_{u}*accur*−

_{f}*accur*= − 13.4%).

_{u}*X*is a dummy variable that is equal to 1 when the target is flanked and 0 otherwise, and β quantifies the independent influence of being flanked versus unflanked on mean accuracy after accounting for the influence of target retinal size. The fitting showed that β = 0.18, permutation test

_{flanked}*p*= 0.20. Hence, there was no significant effect from being flanked versus unflanked for the data below 5° eccentricity. We applied the same analysis to the data above 5° eccentricity to get the baseline mean accuracies and the significance level of the independent effect of being flanked versus unflanked. Baseline accuracies are shown as dashed lines in Figure 8; β = − 0.54, permutation test

*p*< 0.001. Flanked pedestrians versus unflanked pedestrians significantly decreased mean accuracy even after accounting for the influence of target retinal size.

*p*= 0.003). Within 5° eccentricity, there was no significant independent effect from having pedestrian flankers versus car flankers (β = 0.17, permutation test

*p*= 0.30). These results, again, show the target-flanker similarity tuning of crowding in a gender discrimination task.

_{<0.5}= 0.52) and a relatively flat slope over the ratio range above 0.5 (β

_{≥0.5}= 0.01). The difference between the two slopes was consistent with Bouma's rule-of-thumb. To determine the significance level of the slope difference (β

_{<0.5}− β

_{≥0.5}) after accounting for target eccentricity and target retinal size, we added target eccentricity and target retinal size as additional regressors into the clipped line fitting model, and conducted a permutation test where we shuffled the spacing-to-eccentricity ratio values of the trials. The results showed a significant slope change (β

_{<0.5}− β

_{≥0.5}= 0.20, permutation-test

*p*= 0.01; Figure 9B).

*X*is a dummy variable that is equal to 1 when the flanker alignment is radial and 0 when the flanker alignment is tangential,

_{radial}*X*is a dummy variable that is equal to 1 when the target is closer to horizontal meridian than vertical meridian and 0 otherwise, and

_{horizontal}*LPSD*is logrithmic pictorial size difference between target and flanker, and β quantifies the independent influence of being radial versus tangential on mean accuracy after accounting for the influence of target eccentricity, target retinal size, target meridian, and target-flanker depth difference. Results showed that radial versus tangential flanker alignment significantly decreased mean accuracy after accounting for the confounding factors (β = − 0.72, permutation test

*p*= 0.003; Figure 10B). We applied the same calculations to the data with spacing-to-eccentricity ratios above 1 where crowding was unlikely to occur. Results showed that there was no significant independent effect of radial versus tangential flanker alignment after accounting for the confounding factors (β = − 0.26, permutation test

*p*= 0.40). Overall, the result confirmed the presence of a radial-tangential anisotropy in Experiment 2, consistent with the findings in Experiment 1 (see Figure 6).

*accur*−

_{altered}*accur*= − 8.6%; Figure 11A). To determine the significance level of the difference in accuracy between altered and direct saccades after accounting for target eccentricity and target retinal size, we fit the following model to the data:

_{direct}*X*is a dummy variable that is equal to 1 when the trial corresponded to an altered saccade in Experiment 1 and 0 otherwise, and β quantifies the independent influence corresponding to an altered saccade versus to a direct saccade on the mean gender discrimination accuracy after accounting for the confounding influence of target eccentricity and target retinal size. The results showed that altered versus direct saccades significantly decreased mean gender discrimination accuracy after accounting for target eccentricity and target retinal size (β = − 0.31, permutation test

_{altered}*p*= 0.02; Figure 11B).

*Vision Research,*16, 71–78. [CrossRef]

*Vision Research,*45, 1385–1398. [CrossRef]

*Vision Research,*43, 2895–2904. [CrossRef]

*Journal of Vision,*16, 10. [CrossRef]

*Nature,*226, 177–178. [CrossRef]

*IEEE Transactions on Pattern Analysis and Machine Intelligence,*8, 679–698. [CrossRef]

*Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,*4974–4983.

*Vision Research,*41, 1833–1850. [CrossRef]

*Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition,*2016 December, 3213–3223.

*Journal of Vision,*10, 1–17. [CrossRef]

*Journal of Vision,*11, 2. [CrossRef]

*Journal of Vision,*9, 1–15. [CrossRef]

*Science,*142, 979–980. [CrossRef]

*Scientific Reports,*8, 14073. [CrossRef]

*Journal of Experimental Psychology: Human Perception and Performance,*36, 673. [CrossRef]

*Proceedings of the National Academy of Sciences of the United States of America,*114, E3573–E3582. [CrossRef]

*The Journal of Neuroscience: The Official Journal of the Society for Neuroscience,*33, 2927–2933. [CrossRef]

*Proceedings of the IEEE International Conference on Computer Vision,*2961–2969.

*Psychological Research,*71, 646–652. [CrossRef]

*Journal of Vision,*13, 20. [CrossRef]

*Spatial Vision,*8, 255–279. [CrossRef]

*Vision Research,*48, 635–654. [CrossRef]

*Journal of Vision,*2, 167–177.

*Journal of Vision,*2, 140–166.

*Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,*8759–8768.

*Journal of Vision,*7, 1–12. [CrossRef]

*Journal of Vision,*10, 1–14. [CrossRef]

*Journal of Vision,*7, 24. [CrossRef]

*The International Journal of Robotics Research,*36, 3–15. [CrossRef]

*Journal of Vision,*7, 1. [CrossRef]

*Journal of Vision,*12, 13. [CrossRef]

*Current Biology,*28, R127–R133. [CrossRef]

*Human Factors,*57, 701–716. [CrossRef]

*PLoS One,*6, e19796. [CrossRef]

*Nature Neuroscience,*4, 739. [CrossRef]

*Journal of Vision,*4, 12. [CrossRef]

*Nature Neuroscience,*11, 1129–1135. [CrossRef]

*Journal of Vision,*14, 5. [CrossRef]

*Journal of Vision,*9, 1–11.

*Attention, Perception, & Psychophysics,*77, 1252–1262. [CrossRef]

*Psychological Science,*21, 641–644. [CrossRef]

*Journal of Vision,*8, 1–9. [CrossRef]

*Journal of Vision,*11, 13. [CrossRef]

*Attention, Perception, & Psychophysics,*77, 508–519. [CrossRef]

*Vision Research,*32, 1349–1357. [CrossRef]

*Journal of Vision,*12, 6. [CrossRef]

*Vision Research,*15, 1137–1141. [CrossRef]

*Trends in Cognitive Sciences,*15, 160–168. [CrossRef]

*Applied Ergonomics,*65, 316–325. [CrossRef]

*Frontiers in Human Neuroscience,*8, 103. [CrossRef]

*The IEEE Winter Conference on Applications of Computer Vision,*1767–1775.

*Asian Conference on Computer Vision,*658–674. Springer, Cham.

*Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,*2174–2182.

*Journal of Vision,*15, 21. [CrossRef]