Object Detection/Classification

Generic Object Detection

The goal of Generic Object Detection is to detect arbitrary objects, without knowing their specific classes. In general, the detection is based on geometrical aspects.

The segmentation can be done by utilizing stereo information. The first step is to detect and segment the ground plane. Afterwards an algorithm groups coherent objects in 3D.

Further steps could be to classify the objects, or track moving objects between different time steps.

Model based Object Detection

A previously known model can be used for tracking an object's shape through a Bayesian filter.

2D-Detection through Machine Learning

Obect Detection can also be done by utilizing Deep Neural Networks (DNN) or other machine learning algorithms. 

In the following, we trained a DNN to detect different static objects (signs, etc.).

Afterwards, the 2D detections  are tracked through recursive Bayesian filtering in the 3D-space.

These 3D-objects can be used for landmark-based navigation.