Chengke
It is hard to answer this kind of question in the abstract, however matters
to consider:
1. the input traffic flow data - In almost all automated counting systems it
is only the main route structure that is counted. In other words your data
are likely to relate only to the most integrated axial lines in your
dataset. In order to test the hypothesis that there is a relation between
network configuration and traffic flows you would ideally look for data that
are well distributed across the range of both traffic flow and spatial
configuration variables. If your data are restricted on these variables then
your expectation of a correlation at a particular level would drop.
The details of these kind of data are also problematic. What are the nature
of the counting stations and the data you are using? There are several kinds
of traffic counting data that are in common usage, and these tend to be
gathered for quite specific engineering purposes, and are not necessarily
designed for the kind of research question you seem to be addressing. For
example, when data are hand gathered (and these may be the most robust in
terms of vehicle type and numbers) these tend to be derived from studies of
limited duration and tied to very specific local questions - a particular
junction for example that is to be reconfigured. There are always questions
regarding these data since they often indicate that some engineering work
was about to take place designed to change the flows, and full counts after
the changes seldom take place.
If the counts are automated, again these tend also to have problems. The
most common types count axles and often run across all carriageways - these
can be used to estimate actual flows of vehicles, but a number of
assumptions are built into that estimate, including the ratios of different
vehicle types (HGV have more axles than cars.) Data may be derived from
induction loops under the road. Again estimating vehicle numbers and types
is particularly difficult with these data, and especially under congested
conditions. These induction loops are often used to control traffic signals
- they are designed to detect when a queue at the traffic lights has built
up to a particular length. They work by 'polling' whether or not there is a
large lump of metal above them every quarter of a second or so, and the
translation of that to vehicle flows depends on data on traffic speed, car
to car distance, etc. that must be estimated or assumed. This translation
doesn't need to be done for traffic control use, so if your data come from
here you have to check how the base data has been translated into flow data.
2. the axial mapping - am afraid that you have to ensure that underpass and
over pass are disconnected. These lines are often very important in the
structure and to assume they are connected when in fact they are not is a
problem. Unlinking is a relatively simple process in Depthmap, but I agree
it takes time. I also think that you need to check the reasonableness of
your simplifications of complex traffic networks. The key thing is to be
consistent - represent similar situations in similar ways.
3. Attribution of multiple counts to a single line segment. You should be
summing rather than averaging your counts I think. For example you add
counts going in opposite directions on a two way street, rather than
averaging. If your counts are at two sequential locations along the length
of a single line then I would advise putting these in as separate data
points in the stats to start with. You may go on to find that the average
counts down a line correlate bets, but it might not be that. It could for
example be the maximum flow that is the best correlate - this is a research
question.
4. Kowloon can have a lot of congestion at rush hours, and traffic flows can
be severely disrupted by pedestrian flows. In the first instance I would
advise looking at the data only for non-peak times. See if there is a
relationship under free flowing conditions. Does this break down under
congested conditions as previous research would suggest?
5. Size of model - if the counts are just on the main roads then I would
expect you to need a larger axial model, as well as more global radius
measures to best correlate.
Alan
> Deal all,
>
> I have finished drawing axial lines of one district of Hong Kong -
> Kowloon,
> totally 3004 axial lines. Also I pin-pointed all the counting stations on
> those axial lines. I calculated local integration of axial lines and
> compared it with the natural logarithm of the average traffic flow of that
> corresponding axial line. To my disappointment, the correlation is quite
> poor, r square value is only 0.10. I am not sure why the correlation is so
> poor, the possible reasons are listed as follows:
>
> 1.The street network in Hong Kong is very complex, I digitized some road
> lines with only one axial lines, since they represents different lanes for
> different traffic direction. And even, somewhere a cluster of roads are
> represented for only one axial line for simplification.
> 2.Due to lacking of time and manpower, I did not take those overpass and
> underpass roads into consideration when I calculated the connectivity
> value
> of axial lines. When I digitized the axial map, I only referred to the
> digital street and building maps of Hong Kong at hand. Some lines are
> apparently intersected on digital map, but actually they are not, since
> one
> of them is overpass or underpass road.
> 3.When I assigned the traffic flow the axial lines, I first pin-pointed
> counting stations on roads to corresponding axial line, and then averaged
> the traffic flow.
> 4.The spatial configuration (space syntax) is not suitable in Hong Kong to
> predict traffic flow.
>
> The above factors are responsible for the poor correlation from my point
> of
> view. What do you think about this problem? Since Kowloon is only one
> district of Hong Kong, I need to continue with the whole Hong Kong
> territory. If something wrong in above 4 factors, please point it out, so
> I
> can correct it early. Your comments are highly appreciated!
|