Object Detection with Deep Learning

As part of the project AUXILIA and other industry sponsored projects in the field autonomous driving, the professorship is actively researching in the area of object detection using deep learning. Our research focuses on, but is not restricted to application of convolutional neural networks (CNN) on omnidirectional fisheye camera images.

Fig 1: Detection results from the test apartment

'

Fig 2: Detection results from the test apartment

The figures above show two scenarios from our test apartment, where persons are detected in images generated by fisheye cameras mounted on the ceiling. The distortion of objects and the unusual viewpoint poses great challenges on the CNN-based detectors. At the current stage, person can be reliably detected real time in omnidirectional images.

Current research focuses and open topics:

• Expand the CNN object detectors to indoor objects.
• Improve reliability and precision.
• Improve performance on embedded platforms.
• Explore other possibilities other than linear CNNs, such as spherical CNN.
• Explore internal image feature representation and utilization by CNNs on omnidirectional images.
• Train object detectors with synthetic data.