You are on page 1of 16

Mid-scale crowd simulation: a short summary

Jingchao Kong
17.10.2023
Table of Content

• Crowd-driven mid scale crowd simulation


• My further reading
Crowd simulation of T. Feng: an overview

The model consists of the following parts:


1. Layout representation and Optimized object design

Dr. rer. nat. Erika Mustermann (TUM) | kann beliebig erweitert werden | Infos mit Strich trennen 3
Layout Representation
Using graph to represent as:

• Sites, with specific functional type


• Path, with length and width
• Form a graph

4
Optimized object

Consists of two parts, agent cost and prior cost

The agent cost contains 3 parts, mobility, accessibility and coziness.

But in practical it is too costly, instead, they used a regressor as a predictor to compute the
loss.

Dr. rer. nat. Erika Mustermann (TUM) | kann beliebig erweitert werden | Infos mit Strich trennen 5
Agent behavior

• Think, walk and visit

• The time and length of the agent routine is set in advance

• The visitors of the mall is divided into 4 catagories, they experience various time schedule
during visiting. The result is inherited from social science.

Dr. rer. nat. Erika Mustermann (TUM) | kann beliebig erweitert werden | Infos mit Strich trennen 6
Density and speed

The social distance is sampled uniformly in the interval obtained emperically.

The walking speed decreases when the crowd is squeezed.

The methods of computing the density in visiting and walking scenario slightly differ, while
using circumstance sphere to compute the walking and using the area of site to compute
visiting density.

The walking speed model is also obtained emperically. So coarse!

Dr. rer. nat. Erika Mustermann (TUM) | kann beliebig erweitert werden | Infos mit Strich trennen 7
Warning!
1. Usually people do not act completely individually when shopping, for instance, the case of
families, couples, etc.

2. The walking speed decay usually does‘t stay the same in the cases of walking and visiting.

3. Everything is either set priorly or obtained empirically.

Dr. rer. nat. Erika Mustermann (TUM) | kann beliebig erweitert werden | Infos mit Strich trennen 8
Cost functions
1. Mobility: forces the agents to walk smoothly at their natural walking speed
Computed by taking portion of the actual speed and the natural speed

Dr. rer. nat. Erika Mustermann (TUM) | kann beliebig erweitert werden | Infos mit Strich trennen 9
Cost functions
2. accessibility: the agents should be able to visit all the planned sites with shortest path.

Dr. rer. nat. Erika Mustermann (TUM) | kann beliebig erweitert werden | Infos mit Strich trennen 10
Cost functions
3. Coziness: people feel uncomfortable when they are positioned in either too sparse (just
before closing the shopping mall) or too dense crowd (rush hours).

11
Cost functions
3. Coziness: people feel uncomfortable when they are positioned in either too sparse (just
before closing the shopping mall) or too dense crowd (rush hours).

12
Prior costs
• Number of each type of sites

• Sites area compared with the layout area

• Cost weights: all are hyperparams!

13
Features for predictor
1. Geometric features: edge length, edge width, etc.

2. Topological features: Node and edge valence, between and closeness contrality, both
computed from the information of the graph itself.

3. General features: path area ratio, agent density over time.

These features are computed intrisically and used for the classification and prediction of
random forest, whose model is offline trained.
After training the RF, given a new layout input, without computing the loss by agents, we get
the loss by feeding the input into RF, which is an approximated but faster one.

14
How to reduce the loss?
Some actions are proposed:

 Sliding boundaries,
 Changing path width
 Swapping blocks, etc.

Has nothing to do with backpropagation and is definitely not ensured to minimize the current
loss, just in essence an iterative attempt.

15
Warning!
1. The initialization of the layout is not shown. Is there a general initialization method for
different input?

2. Even a non increasing optimization object is not guarenteed, the proposed movements are
all set manually despite of the „temperature settings“ to get rid of the local minima.

16

You might also like