### Abstract

The use of neural networks in the classification of human gait profiles has been in the past few years of interest to many researchers. Human gait has been used as a biometric measure to distinguish between known and unknown persons in cases where other biometric measures such as face are not available. In this research, we analyse simple gait data (cadence, cycle time, stride length, speed) of normal children of age range between 12 and 72 months. We
examine how artificial intelligence, in the form of neural networks, can exploit such data with the objective to correctly identify a child's gender. Data collection and method We used the data set provided by Dr. Sang-hyun Cho for normal children available at http://guardian.curtin.edu.au/cga/data/index.html. We decided to concentrate only on primary gait parameters for simplicity and to examine the role the values of these parameters play in gender recognition. Thus, using the mean and standard deviations provided by Sang-hyun Cho for age, cadence, cycle time, stride length and speed, we generated data for females and males assuming a Gaussian distribution. We have normalised the data such that they would have a zero mean and a unit standard deviation.
We have used two types of neural networks: multi-layer perceptron (MLP) and self-organising maps (SOM). For the MLP, we have experimented with various architectures and parameters of the network to try to achieve the best recognition rate possible. With a two--layer network and
a tanh activation function for the hidden layer and a sigmoid for the output layer, 5 neurons in the hidden layer and a learning rate of 0.9, we achieved a typical testing recognition rate of 57.89% with a mean square error of 0.012 after 17000 epochs of training. An optimum
recognition rate of 63.16% was achieved with 10 hidden nodes and learning rate of 0.5. We used the SOM on the same sets of normalised data as in the previous experiments
with FFBP to examine how the same problem of gender classification, based on simple gait parameters, can be solved by this different type of neural network. A self--organising map does not need target outputs. It was trained only on inputs with the objective to learn their internal
distribution and categorise them. We adopted a hexagonal arrangement as the topology of the neurons, since this is considered to avoid favouring of the diagonals. Then for the ordering phase, during which a rough order of the SOM is performed and the convergence phase which lasts much longer we chose the following values: Ordering phase learning rate=0.9, continuously during the gait cycle. Fig. 2 shows how the SOM projected the 43 curves into
single trajectories. Similar gait patterns are located close to each other. Fig. 3 demonstrates that
a SOM sensitised to ankle abnormality will separate the gait patterns dominated by abnormal ankle curves more clearly. The interpretation of the SOM trajectories of Fig. 4 together with those in Fig. 2 allow the identification of abnormal ankle patterns. The sensitisation factors can be applied to any groups of curves allowing a delicate tuning of the focus of visualisation. The combination of the QE, SOM projections and sensitisation to groups of curves provide a well controlled set of tools which handle the whole of the gait patterns and so allow a direct visualisation of the quality of gait.

Original language | English |
---|---|

Pages (from-to) | 39-40 |

Number of pages | 1 |

Journal | Gait and Posture |

Volume | 18 |

Issue number | Supplement 2 |

Publication status | Published - 2003 |

Event | European Society of Movement Analysis for Adults and Children - Marseille, France Duration: 10 Sep 2003 → … |

### Keywords

- gender recognition
- children
- gait data
- artificial neural networks

## Fingerprint Dive into the research topics of 'Gender recognition of children from gait data using artificial neural networks'. Together they form a unique fingerprint.

## Cite this

Lakany, H., & Spanoudaki, S. (2003). Gender recognition of children from gait data using artificial neural networks.

*Gait and Posture*,*18*(Supplement 2), 39-40.