Table 1: Summary of the amounts of signs found in both 2011 and 2012, together with the amount of signs flagged by our
mutation detection approach. In cases where e.g. the GPS locations deviates significantly, a physically unchanged sign will
be marked as both newly placed (the 2012 sign) and removed (the 2011 sign).
# Ground truth signs Identified sign mutations
# 2011 # 2012 % w.r.t. # 2011 # per mutation category % correct
Total signs 16, 504 16, 548
Unchanged signs 15, 127 91.7% 14, 345 94.8%
Newly placed signs 1, 421 8.6% 2, 541 55.6%
Removed signs 1, 377 8.3% 2, 254 61.1%
Missed newly placed signs 7
In practice, this error source can be neglected, as this
error is caused by the recognition of a sign-like ob-
ject of the same type at about the same position as
the removed sign, which is a very unlikely situation.
Manual correction of errors of Category 4, which is
caused by the detection accuracy of our sign recogni-
tion system, involves browsing through all images to
search for missed signs, which is a time-consuming
procedure, as complete images should be evaluated.
Neglecting this error source may affect the quality of
the updated inventory, because our sign recognition
system currently positions about 93% of the signs. As
browsing through all images is rather inefficient, we
have searched for other ways to retrieve the majority
of the missed signs. Since our recognition system de-
tects about 98.1% of the signs in at least a single im-
age, an alternative would be to evaluate all detections
that are not contained in a positioned 3D sign, such
that the amount of newly placed signs that is missed
is bounded to about 2%. Since this action operates on
detections, this process can be performed efficiently.
Related to this, we have observed that newly missed
signs have a lower probability of being worn due to
aging or being covered by vegetation. However, the
small number of missed signs complicates statistical
quantification of this.
Summarizing, all errors generated by our auto-
mated mutation detection system can be resolved ef-
ficiently by employing limited specific manual inter-
vention, leading to a continuation of the high inven-
tory quality over sequential surveys.
6 CONCLUSIONS
This paper has presented a (semi-)automated ap-
proach for detection of mutations in existing inven-
tories of traffic signs. The system consists of two
stages. The first stage involves the automatic cre-
ation of a new road sign inventory from street-level
images. This process starts by processing all individ-
ual images for sign detection, followed by a multi-
view position estimation process to retrieve the posi-
tions of the detected road signs. Afterwards, all po-
sitioned signs are classified, based on the detections
employed during positioning. The second stage ana-
lyzes the differences between the resulting inventory
and the baseline survey, aiming at the re-identification
of all unchanged signs. This results in the retrieval of
all changed situations, which enables specific manual
validations to attain the target quality.
This system is employed to perform a mutation
scan in a large geographical region over a 1.5-year
time period. The total amount of found changes
equals 16.9% of the amount of baseline signs, which
clearly shows the relevance of actualization. Our sys-
tem marked 94.8% of the unchanged signs accord-
ingly, and retrieved a number of changes equal to 29%
of the amount of baseline signs. We have analyzed the
error categories of our system, and we have discussed
the required manual intervention for resolving them.
These actions operate on a limited set of signs or de-
tections, and thereby allow for preserving the inven-
tory quality level. This approach reduces the required
manual effort with a factor 5, compared to recreat-
ing the inventory from scratch. In addition, this ap-
proach contributes to the feasibility of frequent updat-
ing, which is currently a time-consuming procedure.
REFERENCES
Bonaci, I., Kusalic, I., Kovacek, I., Kalafatic, Z., and
Segvic, S. (2011). Addressing false alarms and local-
ization inaccuracy in traffic sign detection and recog-
nition. In 16th computer vision winter workshop,
pages 1–8.
Comanicu, D. and Meer, P. (2002). Mean shift: A robust
approach toward feature space analysis. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence,
24:603–619.
Creusen, I., Hazelhoff, L., and de With, P. (2012). Color
transformation for improved traffic sign detection. In
Image Processing (ICIP), 2012 19th IEEE Interna-
tional Conference on, pages 461–464.
Creusen, I. M., Wijnhoven, R. G. J., Herbschleb, E., and
de With, P. H. N. (2010). Color exploitation in hog-
based traffic sign detection. In Proc. IEEE Interna-
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
712