Evitech has developed a new version of the function "abandoned objects detection in public transports" based on AI techniques that enable light change filtering, such as the ones that may happen in rolling cars when they run outside.

In the LYNX product, EVITECH had developed a few years ago a function named SAKER whose purpose was to detect people and objects left in an underground train. Light changes in the underground transports are relatively low, compared to outside, so this first version could be realized with classical video analytics techniques, following statistical methods and using the background colors of the train opposed to the foreground to be detected. Geometrical perspective of the scene, zoning, and object size thresholding were also implemented to filter small objects, such as the handkerchief #2. The accuracy of this first version was excellent, with 100% detection, no false alarms, when tested over one full go-and-return course of the train on the line.

 SAKER version 1 : objects & persons left in an underground transport vehicle

When testing this application in outdoor contexts, for export projects, Evitech showed that this kind of approaches could not allow to overcome systematically light variations, water reflections on the ground, and all effects of sun on the ground with a single setting that would cover the whole daylight conditions. An evening setting would be too sensitive by day, and a day setting would not be sufficiently sensitive in the evening or early morning. Light changes onboard are not the ones happening in a scene on a fixed CCTV camera (usually sun and clouds). Train mobility can address paths close to trees and buildings, where alternance of shade and light can be very important and happen quickly, and courses in railways curves can generate sun spots moving quickly with high contrast level on the train ground. Such variations cannot be controlled by a classical approach based software.

SAKER version 2 : from upper left  to down right : detections highlighted in red - laying an object, darkening, sun spots, sunspots translated, door opening (passenger feet + shade), door opened (saturated areas).

By revisiting this application with Artificial Intelligence techniques, Evitech changed its approach and developed a new neural network model for detecting such differences and filtering light changes as well as water on the ground. For doing this, it was necessary to constitute and to annotate by hand an important data base of images. Then a network structure was developed to analyze all characterististics of all train parts and objects (like in VGG layers), extended with a space filtering layer, and producing a probability of foreground at several layers of image resolution. By agregation of pixel based detections, the detection map is reconstituted to provide spots corresponding to "suspicious" objects and persons. Then, the same geometrical and zoning tools as before can be used to propose a list of relevant targets.

Detections here are highlighted in red (the bag is originally white. It appears red, as the persons'feet, due to detection highlighting).

Accuracy of this solution allows to reach outside similar performance as the ones that were reached underground with the initial one, provided a threshold is used on detection size (like for the small spot down right the last image), or very short transient false detections that can be easily filtrered temporally.

In order to limitate the economical impact of DNN based solutions (GPU cards) and to embed the solution onboard, Evitech ported it over a virtual environment that allows the execution of the software over a light GPU unit (such as NVIDIA Jetson nano) and also over pure CPU hardware (CPU based hardware is more accessible, in public transports certified series, than GPU one).