Information Technology Reference
In-Depth Information
spatio-temporal flow data attributes. In order to confirm the results obtained with this
data set, the next step of this research will evaluate the system by using a larger num-
ber of both accident and non-accident cases, and also by varying the percentage of
accident/non accident cases in the case base. Currently, the case base consists of 39
accident and 32 non-accident cases, so for each case, the probability that retrieval is
successful is 54% for accident cases and 44% for non-accident cases. Increasing the
percentage of non-accident cases in the case base might reduce the success rate, how-
ever considering that currently the success rate for both accident and non-accident
retrievals is comparable, we expect that the system performance is relatively stable to
changes in the contents of the case base. However, this will be confirmed using expe-
riments. Currently, most attributes describe the traffic flow. However, it is likely that
accidents occur also due to other conditions, such as the weather, the light conditions
or visibility and human behaviour. We will use another accident data set consisting of
police accident records, which contains a wealth of additional information, including
weather conditions, types of accidents and the severity of an accident that could be
important in determining accidents [19]. It is difficult to directly measure the contri-
bution of individual human error resulting from their state of mind to accidents
though without doubt they play a large role. However, human error is expressed par-
tially in their driving and therefore in the flow data, for instance, speeding or insuffi-
cient distance between vehicles. The similarity between cases with respect to the
normalised spatio-temporal flow attributes is computed using the Euclidian distance.
We are currently working on implementing a more accurate similarity measure based
on the Mahalanobis distance, which takes into account the correlation between
attributes. Another important consideration is case attribute selection and weighting.
Not all attributes are significant with respect to the desired solution of a target case
and the contribution of each attribute to the similarity between cases can differ, too. A
possible method of attribute weighting would be to use local weights, which vary
based on the attribute values in the target case and take into account the relation be-
tween attributes [20]. Currently, the spatio-temporal attributes are computed by
matching corresponding points in time and space. However, often spatio-temporal
patterns can be similar even if one pattern is stretched, contracted or of a different
length than the other pattern. This is possible using dynamic time warping [21]. It
would also be interesting to evaluate how the accident prediction system using
CBR and spatio-temporal data compares to other accident prediction systems in the
literature [5, 9].
The next stage in the system design involves finding the threshold above which
cases similar to the target case are deemed to be good predictors of the outcome of the
target case. Given a target case, if the similarity between the retrieved case and the
target case is larger than the threshold, the CBR system will predict that the outcome
of the target case is similar to the outcome of the retrieved case. The advantage of
CBR is that it inherently detects traffic flow patterns that are indicative of accidents.
However, the accident cases in the case base (and their similarity) can be studied to
identify a set of conditions that might lead to accidents. These could be used by traffic
controllers or designers a priori to reduce the likelihood of accidents. The aim of the
final system is to automate this process so that flow data is analysed continuously
