I'm not sure there's totally indisputable evidence on this, but I found a newer (2017) paper by Bayet et al. with infant subjects, but with fear rather than anger as the emotion tested (vs. happy), and (more importantly) using a different methodology (and IMHO better) because instead of different faces (in a crowd), they used the same face but mixed with noise, e.g. for fear [but for illustrative purposes only, see blow why]

They actually tried to determine which facial features enabled easier detection, eyes or mouth by mixing noise preferentially in that region (the resutling images were validated with SSIM, so there were four series of stimuli prestented (whith varying level of noise in each series).
Each of the face stimuli were presented in differents tests to the infant subjects; each test consisted of a noise-mixed face vs pure noise (two images on screen) and a camera was used to record the infant's reactions. The film was then analyzed to determine several measures of reaction, which were in turn used to calculate measures of face detection.
Two different measures of face detection were ultimately derived: A simple one based on percentages of total looking time [PTLT] and a more complex psychometric measure, the methodology of which is pretty involved (and thus perhaps a weak point of this paper):
A multivariate measure of face detection was derived by classifying
trials as ‘face is on the left’ or ‘face is on the right’. The
rationale for this metric is similar to the idea of ‘double psychophysics’
[48]: if one can reliably guess on which side of the screen
the face was presented by looking at the infant’s behaviour, then
it can be inferred that the infant is discriminating between the
presence of a face or noise. We implemented this idea computationally
using MVPA [42] to locate the side of presentation of the
face (left or right) on each trial based on (i) PTLT to the left,
(ii) number of looks to the left, (iii) number of looks to the right,
(iv) duration of first look to the left, (v) duration of first look to
the right, (vi) median duration of looks to the left, (vii) median
duration of looks to the right, and (viii) direction of first look
(left or right). Durations were log-transformed [56]. Continuous
measures were z-scored within-subject. Measures were chosen
a priori given the visual preference of infants for faces [57].
PTLT to the right is equal to 100% minus PTLT to the left, and
thus did not need to be included. Trials from all participants
were pooled to maximize the number of training examples, and
a logistic regression algorithm (a common classifier for MVPA)
was repeatedly trained on all trials except one and tested on the
trial that was left-out (leave-one-out cross-validation). Forward
sequential feature selection was implemented inside each crossvalidation
loop (see electronic supplementary material, table S2
for results on the full dataset). This procedure led to locating
the face side for each trial in a way that reflects generalization.
We used logistic regression because it provides log-odds, a
direct, criterion-free, continuous measure of evidence for each
response (‘face is on the left’ versus ‘face is on the right’)—as
opposed to accuracy, a binary measure dependent on a decision
criterion. Raw evidence (log-odds for the right versus left side)
was pooled to derive correct evidence (log-odds for the correct
versus incorrect side) as a multivariate measure of face versus
noise discrimination.
The overall results of their analysis look like this (one row for each detection measure):

Their interpretation of results:
Psychometric curves of the PTLT to the face side revealed
a significantly lower threshold for the Fearful eye+ face condition
(44.41 ± 1.98% face signal; [...]; figure 3a) than for the Happy eye- face
condition (difference: 5.20 +/- 2.62% face signal, 95% CI
[0.001 0.103]), but not the other conditions ([...] figure 3a), across all age
groups. [...]
A similar result was found when applying psychometric
curve modelling to the correct multivariate face versus
noise discrimination evidence; the face detection threshold
for the Fearful eye+ condition (44.07 ± 2.14% face signal;
[...] figure 3c) was
significantly lower than the detection threshold for the
Happy eye- condition (increase in threshold: 7.90 ± 2.52%
face signal, 95% CI [0.030 0.128]) but not the other conditions
(Wald confidence intervals, alpha = 5%; [...]; figure 3c)
Similar models were used to estimate the difference in
threshold between the Fearful eye- condition and other conditions,
or between the Happy eye+ condition and other
conditions (electronic supplementary material, tables S5–S8).
Results are summarized in figure 4. Overall, psychometric
curve modelling of infant looking data revealed face detection
thresholds at about 44% signal, with an increase of
about 5% signal in threshold for Happy eye- condition
compared to the Fearful eye+ condition, and intermediate
thresholds for Happy eye+ and Fearful eye- conditions
depending on whether PTLT alone (figure 4a) or correct
multivariate discrimination evidence (figure 4b) was used
as a measure of face versus noise detection.
To clarify whether these differences in detection
thresholds reflected a main effect of facial expression, a
main effect of eye visibility, or an interaction between the
two, and to test for an effect of age, we conducted further
analyses focused on the linear portion of the psychometric
curves corresponding to trials around the fitted detection
thresholds (40–50% signal). [...; figure 3b,c,e,f I'm omitting all the chi-square stats]
Results were mixed when considering PTLTs for the
face side alone, as the effect of face emotion was restricted
to 3.5-month-olds on this measure. However, correct discrimination
evidence, a more comprehensive multivariate
measure inclusive of PTLTs and other aspects of looking
behaviour (e.g. first look) revealed a detection advantage for
fearful faces compared to happy faces across all age groups
(i.e. it did not significantly interact with age).

Of course this doesn't answer the question whether the effect holds in a crowd (or any non-laboratory setting, or if applies to anger etc.) The authors themselves say:
Future research should determine
whether the readiness to detect fearful faces (compared to
happy faces) in infancy generalizes to naturalistic settings
beyond the laboratory. Comparing detection of fearful
versus angry versus sad faces will also clarify whether the
effect applies just to threat-relevant stimuli (anger and fear)
or more generally to negative novel expressions (fear,
anger, and sadness).