Selected Projects

Scroll to top

Flexible Stereo: Long-Range Depth Estimation

Keywords: Depth Estimation, Visual-Inertial State Estimation, EKF
Long-range depth estimation is crucial for fast-flying UAVs. Depth cameras, radars and LiDARs are either too heavy, bulky, expensive and/or do not meet the range and resolution requirements. For fixed-wing UAVs it is tempting to mount cameras on the wings for a maximum stereo baseline and minimized depth uncertainty (cf. Fig. 1 and Fig. 2). However, as shown in Video 1, the wings are non-rigid.

Video 1: Video shows the flapping wing of the UAV AtlantikSolar.
Fig. 1: Fixed-wing UAV Techpod with indicated stereo baselines
Fig. 2: Relation between stereo baseline and depth uncertainty
A camera mounted inside the fuselage is looking at the wing-tip cameras that have apriltags mounted.

GIF 1: On-ground
GIF 2: During flight
Fig. 3: Overlay
(on-ground vs. during flight)

Inertial measurements, visual measurements and relative pose priors are fused in an EKF. Given the estimated relative poses, standard stereo blockmatching is used to compute the depth images and 3D point cloud.

Fig. 4: Image streams
Video 2: Comparison of depth estimates to offline SfM (Colmap) solution
GIF 3: Rectified stereo images with epipolar lines (top). Depth images (bottom).

More information can be found in Paper [1], [2].

Scroll to top

Deep Learning Based Human Detection

Keywords: Object Detection/Tracking/Localization
For a project with the Swiss-Air Rescue Organization Rega , I designed and implemented the hardware and software for a human detection, tracking, and localization framework. A large amount of realistic data was collected with a drone equipped with an optical-infrared camera rig. The annotated data was then used to train state-of-the-art detectors like YOLO and retinanet.

Fig. 1: Optical-infrared camera rig
Video 1: Detections (retinanet)
Video 2: Tracking

The project has been featured in several news articles and TV reports.

Fig. 2: System overview
Fig. 3: Sensor integration
Video 3: News report (SRF)

More information can be found in the preprint [1].