On one of the first worldwide industrial players ' site, security at work was challenged in case of processes mixing people and machines. For reducing the risks of work accidents, and for optimizing the processes, it was decided to automate these processes with the use of Artifical Intelligence (AI), returning situation information from image analysis to process automata.
Process automation here includes the dynamic signalling on site (lights and displays), automation of access doors and machines, surveillance and control of operations, with the detection of forbidden situations. Surveillance concerns around twenty process platforms or more, each of them being supervised by several cameras covering the set of all process points of view. Process as a whole is only understandable in a three dimensions approach. Typical situations are illustrated in the image here (Picture taken from the Internet, not from the site for confidentiality purpose).
These controls, for the image analysis part, require a learning of all objects shapes, work accessories, vehicles, various means for transporting loads, and tools, to elaborate a diagnosis of situation and send information and alarms to a central process management automata. This automata optimises steps and can stop all activity according to the situation. System criticity is very high, since it is involved in the security of a process that can possibly generate accidents, injuries, which can be serious or catastrophic. One must target an analysis reliability at 99.99 % for accident prevention, and that will optimize the process in 98% of other cases (by quick execution and succession of expected tasks, without dead time).
A database made of quantities of images allows the automatic learning of objects shapes in all positions and situations by training a Deep Neural Nework constituted for this project. For reaching very high performance of recognition that are compatible with safety requirements of the system, it is compulsory to prepare dedicated neural networks that are trained on images taken in the actual context (situation, backgrounds, lightings, cameras, ...), and to complete the approach by using complementary algorithms which reinforce the automatic analysis done by these networks : color controls, sizes, shapes, number, speed, position...), It is the role of our Jaguar software where all these functions are integrated, with the help of its alarm engine allowing to specify complex conditions based on supervised situations.
Jaguar's reliability and it low false alarms rate have been at the root of its selection for this project.
"Jaguar detects in very difficult situations, we know it for years of site protection from our past experience. But here, it's a new activity for our group which was formerly focused on sensitive site protection and fire detection : it's a development axis that we want to promote in the future", the system installer says, approved by the Jaguar development manager at Evitech.
