Opendata, web and dolomites

Report

Teaser, summary, work performed and final results

Periodic Reporting for period 1 - Gait Biometrics 3 (Main goal of the project is to create a prototype of the software, which will be able to identify peoplejust based on the way how they walk.)

Teaser

Our research performed during the first phase of SME instruments grant program revealed that our proposed solution of Gait recognition is based on viable concept. We have made significant progress in this area and although there are several technical issues to resolve, we...

Summary

Our research performed during the first phase of SME instruments grant program revealed that our proposed solution of Gait recognition is based on viable concept. We have made significant progress in this area and although there are several technical issues to resolve, we proved that our originally intended approach to the Gait recognition is working.
For example, we developed an environment for testing of various Gait recognition techniques that allowed us to validate individual methods and prove the concept prior to developing final gait recognition application. This was very efficient approach that enabled us to test and reject methods that were not good enough. The testing was done on three different databases that enabled quick achieving of solid results.
Each database had different number and quality of samples:

• Database contained walkers captured using special sensors offering very precise 3D record of the walk and used samples of 74 people.
• Second database contained walkers recorded in laboratory conditions with good and stable light, still camera and the same angle conditions. It contained 174 samples captured by common CMOS camera.
• The third database contained records taken in the normal outside weather conditions. The camera took long footage of video of hundreds of walking people.

We considered about 8000 different parameters obtained from the human body during walk. We call these parameters “Features”. Features were combined to a vector and as such, each Feature contributed to the final capability of the vector to recognize people. Our approach finally pointed out to a several dozen Features that returned the best results. This allowed us to recognize people with precision of:

• FAR - 0.8% (chance the method marks one as somebody else)
• FRR - 8.7% (chance the method marks a person as new, despite he/she was already recognized before)

Further investigation of the patents that were close to our Gait recognition method led us to the 51 patents selected for closer look. Together with IPR experts we investigated its relevance towards our approach and clarified there were no conflicts.
Evaluation of the business case showed significant interest from the commercial companies as well as from government bodies such as military and police. We believe that the potential for use of our system would justify the effort to develop it. Taking overall results of our work so far we concluded to continue with the development of the Gait recognition.

Work performed

Work performed during the reporting period and main results achieved so far
As during Phase 1 SME is supposed to be focused on feasibility study most of our work was connected to theoretical analysis and their verification on sample’s databases.
Main achievements of our work were:
• We improved and finished debugging tool, which allows us to debug and visualize results from our gait biometrics libraries. This was very important for understanding of current status and making the right decisions on how to continue in the development.
• We built testing databases. These testing databases are very important for us as we can validate our approaches on them. We will continue in collecting more samples from our outdoor installation to end up with large testing database, where we will be able to proof potential customers the quality of the method.
• So far we identified and tested around 8000 features describing the human walk. We evaluated many combinations of features (feature vectors) and created their chart based on their ability to distinguish people.
• We tested most of known approaches how one can separate object (silhouette) in the video. We realized that none of known approaches are fully suitable for our needs. We will have to invent our own method in the future. The method will be designed to identify just parts of the silhouette, which are important for our Feature recognition method. There we will have high demands on precision and speed. We made big progress in identification which parts of the human silhouette are worth considering for people recognition. Also what characteristics generated from these parts have potential to be unique enough for particular person. We spend thousands of hours on this task and now we have great know-how for further development and finishing the task.
• We verified the commercial potential by carrying out detailed market research and competitor analysis. Here we used services of Antonín Hamřík, our external expert with many years’ experience in this field.
• We found one reference customer – ELTODO Group, which offered us their premises for performing tests and validating our concept. The customer is also prepared to take advantage of our results commercially.
• We performed detailed IPR analysis with result that our approach doesn’t infringe any third party.

Final results

Our current results made us believe that the project offers promising potential and commercial viability. Our current results (0.8% in FAR and 8.7% in FRR) were presented to our reference customer (ELTODO Group) and we received positive feedback regarding our further cooperation. They see viability of the project mainly from the perspective of CCTV systems maintainer in Prague so better results in FAR and FRR parameters were requested. Target was defined as <1% for both attributes proved on 2000 samples taken outdoors. We are fully convinced that this target is achievable in a perspective of 2 years.

As this is quite far in the future we defined less strict criteria as first milestone on our path, which would be acceptable for business sector, where precision of human recognition is not so sensitive. For this milestone we defined FAR and FRR at the level of <5% (proved again on 2000 samples). This target is realistic in one year.

The most important obstacle to achieve is the issue with separating human silhouette from the background and recognizing Features. This shall be our primary focus for upcoming months.

We have also several other ideas about improving the method and increasing system reliability. We are planning to test those and check if they can improve quality of our results with less demand on computer performance and time needed for analysis.

We are convinced (and we have an endorsement on this from ELTODO Group and Professor Jiří Straus - distinguished expert in criminalistics and forensic biomechanics) that if we manage to reach our targets in system reliability, we shall have great tool for both government and business sector.
Governmental sector can benefit from the proposed solution in the following domains:

• Police: Many times police have high quality records of suspects that are hiding their face. Our system can help recognize such people and trace them as they move in front of cameras. Going further to the future, if a suspect has a record of unique walk pattern in the police database, he/she can be recognized while the system processes video from this new case.
• Customs office: People that committed crime in the past can be recognized at the airports and special attention can be put to them, regardless if they changed their name, outfit or passport.
• Military: The system can be used for protection of restrictive areas as a supplement to authentication systems.

In the business sector, the use can be expected everywhere where recognition of people is important from the security or from client service perspective. We have already started the mapping of the situation in the Czech Republic, together with Antonín Hamřík – a consultant for larger companies and business expert in the field of information technology applications. There is a reasonable and still growing number of cameras (about 30,000 in 2015) and where we intend to acquire our first reference customers:

• Banking sector: There are about 45 different banks with more than 2300 branch offices that can find value in the early recognition of visitors entering the branch.
• Shopping centres: There are 139 big centres and most of them have their own security department. Promising category of customers are also supermarkets (around 700).
• Gas stations: There are more than 3,648 public gas stations, most of which have cameras.
• Any restricted or guarded areas: There are around 830,000 of them, from which 90% are too small for camera installations. Around 20,000 guarded parking lots can be considered as our target. These are parking areas where guards can receive an early warning about return of previously misbehaving people.

Numbers and facts discovered on Czech market will be compared with European markets and used for forming the expansion strategy in the future. More details will be provided in Final report document.