A team of British and Spanish developers have proposed a method for recognizing a person by his gait. A neural network based on the deep residual learning method allows a person to be recognized by the spatial and temporal characteristics of his footprint with almost one hundred percent accuracy. This is reported in an article published in IEEE Transactions on Pattern Analysis and Machine Intelligence.
Traditionally, to authorize and restrict access, data or means are used that are available to a narrow circle of people: keys, passwords, tokens or special cards. However, the password can be guessed, the card can be stolen, and recently there have been methods of forging biometric data that are individual for each specific person: a fingerprint, retina, and even a face. Therefore, there is a need for more effective protection - in particular, effective methods of granting access to only one specific person are required.
One type of biometric that can be used as an identifier is the individual characteristics of the human gait. Such characteristics are divided into spatial and temporal: the first includes measurements of the points of contact of the foot with the support (foot turn, step length and its base, that is, the position of the foot surface), and the second refers to the duration of various (support and motor) phases of the step. A large number of factors affecting the individuality of the gait reduces the likelihood of its copying to a minimum; however, in a real situation, such recognition can be complicated by external factors. For example, in order for a computer to assess gait, computer vision technology can be used, but it will be necessary to make sure that the observed object is in full visibility,which is impossible to provide in low light or crowded conditions.
Scientists led by Omar Costilla-Reyes from the University of Manchester suggested using foot images for recognition by gait. To develop such a method, they collected a database of more than 20 thousand images of footprints of 120 people, obtained using 88 piezoelectric sensors that calculate the magnitude of pressure, on the basis of which heat maps of its distribution are created depending on the phase of the step. Volunteers participating in the data collection were asked to wear any comfortable shoes and demonstrate their natural gait.
Sample of raw (top row) and processed (top row) data for the tracks of two (ab and cd) people from the sample. Costilla-Reyes et al. / IEEE Transactions on Pattern Analysis and Machine Intelligence.
To train the recognition system using the collected data, scientists trained a deep neural network based on the residual learning method, which makes it easier to train a model with a large number of layers (with greater depth), which are often necessary for efficient recognition of images with a large number of parameters. Recently, using this training method, they learned to predict the behavior of a dog by its gait.
The model was tested on three datasets of different sizes, corresponding to different recognition situations: checking at the airport, checking at the workplace and at home. The recognition efficiency depending on the dataset (from the smallest recognition at the airport to the data collected "at home") ranged from 92.9 to 99.3 percent.
The authors note that, as with most similar models, the effectiveness of their recognition system directly depends on the collected dataset: it can only recognize those people about whom it has data. However, collecting data with floor sensors and third-party cameras is a much more real-life task than collecting fingerprints. It is not yet clear how the developed model will cope with possible temporary gait anomalies, for example, after a fracture or sprain.
Promotional video:
Elizaveta Ivtushok