Evaluated Model of Pedestrian Movement Based on Space Syntax, Performance Measures and Artificial Neural Nets

To understand the generating causes of pedestrian movement is very important for urban planning tasks, because it is possible to infer if attitudes taken in the conception and maintenance of the spaces are in fact contributing to the social dynamics. However, to determine pedestrian flows is a difficult task due to the complexity of people movement. A way to outline this problem is through the creation of models which associate the attributes and their relationships directly to the studied phenomena. The model used here uses two kinds of variables: the configurational ones obtained through the axial map of the city where the study area is located, and the performance measures obtained through the physical evaluation of the attributes of the sidewalks of the studied area. The output of the model is the mean pedestrian rate of the area. The Syntax Space theory is useful for understanding the phenomenon because of the way it deals with space interaction. However, though it is able in predicting part of the movement, we do not find significant correlations when the measure of intelligibility is low. Pedestrian flows are a complex phenomenon and, per se, cannot be understood through linear relationships among any couple of variables, being them spatial or not. In this paper it is argued that the space syntax theory and measures explain the pedestrian movement as a phenomenon emerged from society, but the linear approach is not capable of explaining their relationships. The presented model uses Artificial Neural Nets (ANN), a parallel processing tool with the capacity of working through examples, learning, generalizing and abstracting the variables information and their connections. The implementation of these kinds of models evolves from 'black boxes' to models that can be 'disassembled' and evaluated inside of its logical structure. The ANN uses two groups of data: one for training nets and the other one for validating the network. Thus, the performance of the ANN can be tested with unknown data. The results produced so far have shown that ANN can learn the main features of the data sets with an accuracy of more than 90% of correlation coefficient and with an average error smaller than 0.02. It must be said that the research work targets to spread the samples from different configurational realities and expand the data bank on measured movement in order to improve the accuracy of the model, such as being done lately.