Machine Vision for Safety Applications

Innovative Detection Algorithms

Identifying threatening situations involving individuals is the challenging domain of individual safety from falls off the perimeter of a facility or a ship. Our unique offering consists in the development of a data fusion middleware that builds on an innovative set of image recognition and motion detection algorithms, aimed at lowering false alarms and achieving accurate detection of real (human-shaped) targets.

We selected a combination of non-determinism and voting among the classifiers, whereby many different weighted mean values are computed, whose weights are produced by a probability function, and the sum of the medians that exceeds the threshold is computed, leading to a possible alert condition. Our promise is that the system operates better than any one of the individual classifiers, and additionally minimizes false alarms.

Adapting to the Environment

The obvious problem in such applications is the lack of training data representative of the multitude of conditions observed. Additionally, the operation of any detection system at sea is dramatically affected by the weather conditions (rain, mist, fog, hail), the variable background imagery and the movement of the platform itself. In our implementation, the probability function that produces the weights can adapt to changes in the environment (e.g. rain that negatively affects the visibility of the camera), so that other, more reliable metrics get higher importance.

Future Applications

The applications envisaged for the future include perimeter protection – including border protection, facility intrusion detection and protection against maritime piracy in international waters.

The developed system may also be adapted for use by semi-autonomous vehicles for enhancing the safety of pedestrians, bikers and animals including pets, in a way similar to the Autonomous Pedestrian Detection systems currently evaluated by EuroNCAP.