Using AI and Computer Vision to detect Dangerous Objects under the car using Neural Network
Vostok AI was approached by system integrator of security solutions. They wanted to develop easy-to-use but advanced system that allows to quickly scan and analyze if any dangerous objects or armors underneath the car. The solution is planned to be used at military sites, border posts and places with high safety standards.
Challenges with Existing Solution.
The existing examination process was very old school and involved use of mirror and portable photo camera to do car inspection. This process took long time, involved three people and was prone to human error (which could be fatal), thus should be upgraded. The customer wanted to both save time per each checkup, as well as minimize the risks of human error.
Solution.
In theory the project implementation was easy (see image on the left), in reality it took some extra effort to deliver it. The process looks next way:
Firstly, we provided ML-audit service to identify potential use case;
As a result we found a scientific article on a similar topic;
We reproduced the article to our data;
Created custom visualization interface to display the results for operator.
Project Results.
The project was delivered in few stages. Due to nature of operations we had very small data set of images (350 in total) and manually marked the anomalies on the images. Based on this data set we cut photos into 512-pixel squares and trained our model on them. First iteration of model could reach 75% accuracy in identifying abnormal (potentially dangerous) objects. Second iteration could get up to 83% accuracy. Now we are working on 3rd iteration to reach accuracy close to 90% which will make it possible to use this algorithm for on-site inspection in a production mode.