The AI

The main feature of the GUIDE-Walk is the artificial intelligence for object detection, which is supplemented by the other sensors. The AI recognizes 10 different classes in the images of the camera (pedestrians, cars, bicycles, motorcycles, buses, chairs, benches, bins, red and green pedestrian lights) and sorts them according to their position and importance. If something comes too close, the device can warn the wearer accordingly by means of a voice output.

I decided on using the modern object detection network MobileNetV2, as it has the highest accuracy of comparable lightweight models. This is combined with an SSD addition, which goes back to a paper by Wei Liu and which makes it possible to process images much faster and more efficiently.

Such an SSD network consists of a support network and the SSD add-on. The support network generates a feature map from the input, which is divided into a grid by the SSD add-on. For each tile, a certain number of bounding boxes with fixed aspect ratios is generated, which are then checked for accuracy. After further filtering by non-maxima suppression and the removal of duplicates, the best 100 detections are output.

Demonstration: