by Shreyas Kar
This is unbelievable but true that It is estimated that 8 million children go missing around the world every year out of which 800,000 are in the United States alone. The whole world is plagued by child trafficking. Children are taken without their consent and then later sold or used at different places in the world. It is estimated that human traffickers are making $150 billion profit out of this horrendous trade and most of these money fund terrorism and other bad stuff.
The real story that inspired me to do this project happened to me 12 years back. I was lost in a crowded and huge shopping mall. My parents notified the mall authority immediately and gave a picture of mine. Even there are CCTV cameras every nook and corner of the mall, it took a solid 1 hour for me to get reunited with my parents. In retrospective, if this app would have been used by the shopping mall, an immediate match would have been found with my picture to the nearest camera.
One needs to add a photo or just take a picture and feed it to this tool. The tool then processes the image and tries to find similarity with a face in a database of photos and videos. If it matches above a certain similarity level configured in the system will output a match with details about the person it matched. And this happens in real time.
This is the architecture I used to build this project. This is the software program which takes photo as input and another photo or video as other input which can come from the database too. For facial recognition I implemented a Convolutional Neural Network (CNN) to recognize a face against an old face ad trained it with 10,000 images found on internet, namely Flickr-Faces-HQ Dataset (FFHQ). I wrote the program in Java and another app in Swift for iPhone The output is the similarity of the match.
Post-coding I started testing the tool for similarity levels so that the correct level can be configured for the tool to use for right matches. I tested the tool with a current photo of five family members with their old photos at different ages. The data shows that the similarity levels began to level off below 46% age difference level for fully grown adults of the target image. However, people who have not fully grown will level off below 80 percentage difference level. This means that the AI program works better for children than adults. This is because toddler pictures can be more easily compared to adolescents than adults. Now the similarity level can be determined. To do so the similarity level of all the cases where the AI tool works will be listed, then the data will be ordered in ascending order and the 5th percentile value will be taken as the similarity threshold level. When the test was done an 89.52% similarity threshold was determined. A person’s face changes drastically up until around when the person is 14-16 years of age. Thus the similarity level of the person when he is 81 years old will be similar to when he is 35 and when he is 58 but different and a lot smaller than when he was 12. Thus the AI tool works better if the source image is of a child below the age 16 than if the source image is of adults (above the age of 16).
I built an iOS app to get the POC done. These are the screens and flow.
This software can be used by law enforcement agencies in multiple ways. If a home security camera captures a trespasser, it can be sent to law enforcement agency and this tool will help to search with an existing criminal database and can potentially catch the person before committing any further crime. Remember the case that inspired my project. Yes - this tool is used by shopping malls, stadiums, wedding parties etc, wherever there is a big crowd, to match a lost person with live CCTV footage and can be reunited almost instantaneously. If a child is found, he or she can be matched with missing children database and be reunited with family. This tool also can detect and rescue missing and exploited children who are kidnapped. By matching a photo with the live video feed from public places like train stations, airports, Bus stands and alert law enforcement.
