Car, Bicycle and Pedestrian Detection in Adverse Weather Conditions using Lidar and Thermal Imaging

Johannsen, Bjarne Peter

2. term

Vision, Graphics and Interactive Systems, Master

2021

2021-05-26

64

Dette speciale undersøger, hvordan man kan kombinere data fra to sensorer (fjern-infrarøde termiske kameraer og lidar, laserbaseret afstandsmåling) for bedre at opfatte omgivelser i dårligt vejr. Arbejdet adresserer også et kendt problem i termisk billeddannelse: at de datasæt, som dybe neurale netværk ofte fortrænes på, ikke svarer til termiske billeder. For at mindske denne domæneforskel anvendes kunstige (syntetiske) termiske billeder, så underrepræsenterede objektklasser kan trænes bedre. Metoden evalueres i let tåge, tæt tåge og sne, både i dagslys og om natten. Resultaterne viser, at fugt i atmosfæren giver store udfordringer for infrarøde og fjern-infrarøde bølgelængder, og at både lidar og termisk billeddannelse får forringet ydeevne under disse forhold.

This thesis explores how to combine data from two sensors (far-infrared thermal cameras and lidar, laser-based distance sensing) to improve perception in adverse weather. It also addresses a known issue in thermal imaging: deep neural networks are often pre-trained on datasets that do not match thermal imagery, creating a large domain gap. To reduce this gap, artificial (synthetic) thermal images are used to improve training for underrepresented object classes. The approach is evaluated in light fog, dense fog, and snow, during both day and night. Findings show that atmospheric humidity poses strong challenges for infrared and far-infrared wavelengths, and that both lidar and thermal imaging experience degraded performance under these conditions.

[This summary has been rewritten with the help of AI based on the project's original abstract]

Thermal ; Lidar ; Object Detection ; Adverse Weather

Download
View record in AAU Student Projects