TRAFFIC ACCIDENT CLASSIFICATION WITH NEURAL NETWORKS WITH PSO ALGORITHM



TRAFFIC ACCIDENT CLASSIFICATION WITH NEURAL NETWORKS WITH PSO ALGORITHM

Authors

Jose Fierro Moya, PUCV, Cecilia Montt Veas, PUCV, Nibaldo Rodriguez Agurto, PUCV

Description

The ninth cause of death worldwide belongs to people in traffic routes. It is estimated that by the year 2030, this will rise to the fifth position,in Chile, the situation is not very different. We study how to classify accidents, in order to better understand the consequences of certain traffic conditions to raise awareness and reduce the social and private costs, we will study the artificial neural networks and evolutionary algorithms to achieve a classification so as to the type of accident, in other words, to determine if a person is unharmed or injured in a accident, given the input attributes of the network. It develop a model for classifying the severity of those injuries in traffic accidents, using neural networks with optimization algorithms of swarm of particles. The activation functions for the nodes of the output layer were the sigmoidal function in the case of an output neuron, which represents the conditions of the person after the accident, uninjured or injured.

Abstract

The ninth cause of death worldwide belongs to people in traffic routes. It is estimated that by the year 2030, this will rise to the fifth position, WHO (2009). 1.44 million teenagers die each year due to traffic accidents, according to the UNICEF report, April 2012.

In Chile, the situation is not very different. Traffic accidents have been positioned as a true social epidemic, reaching alarming and complex figures to address. According to the National Traffic Safety (CONASET), only in 2011 there were a total of 62,834 accidents, the collision being the most recurrent type of incident. The region of Valparaiso in Chile, analyzed in this paper, is the second region with the highest rate of traffic accident nationwide and 151 lives were lost, according to CONASET (2012). The 2.5% of accidents are fatal, making it necessary to classify them as acceptable.

Given this, we study how to classify accidents, in order to better understand the consequences of certain traffic conditions to raise awareness and reduce the social and private costs. As mentioned above, in this paper, using data from the V Region of Chile, we will study the artificial neural networks and evolutionary algorithms to achieve a classification so as to the type of accident, in other words, to determine if a person is unharmed or injured in a traffic accident, given the input attributes of the network, which are, for example: commune, age, cause, driver, passenger and pedestrian crash, among others.

Currently, while there are traffic accidents investigations conducted by various techniques (Bayesian networks, MLP and SVM etc.), the use of artificial neural networks in the classification of traffic accidents is reduced, and its approach will depend on how the researcher wishes to build the model, which is why this research focuses on the use of artificial neural networks for the classification of traffic accidents.

Therefore, we will develop a model for classifying the severity of those injuries in traffic accidents, using neural networks with optimization algorithms of swarm of particles.

The artificial neural networks require training their weights to achieve a good classification, which is why we have selected the PSO evolutionary algorithm for this task. While there are other methods such as Back Propagation, PSO was chosen for its simplicity and efficiency (Mohaghegi S. et al 2011).

The activation functions for the nodes of the hidden layer were the sigmoidal function or the hyperbolic tangent function. The activation functions for the nodes of the output layer were the sigmoidal function or the hyperbolic tangent function in the case of an output neuron, which represents the conditions of the person after the accident, uninjured or injured. A training exercise was conducted by PSO variants; these are LDWPSO, QPSO, and LDWQPSO. The amount of particles used for PSO variants was 100. Regarding the performance function used by these variants of OSP, these were: error based on the accuracy and cross entropy.

Finally, we compared the results obtained from the average of several samples with QPSO with related works that do not use an average of samples, but the best for a particular sample, therefore, differences could be even higher with the model obtained in this work. For greater accuracy, the best result was obtained by LSSVM FIL trained with PSO, followed by the sigmoidal hyperbolic tangent artificial neural network trained with QPSO and Bayesian networks.
This study also analyzes the network with three inputs and found that better results were obtained; thus, future research will analyze, for example: the percentage of uninjured or injured in an age group for a specific cause and a type of accident. One could say, for example, that people between 15-25 years in a certain percentage are injured for not respecting traffic signals, causing a collision.

Publisher

Association for European Transport