Seite - 98 - in Joint Austrian Computer Vision and Robotics Workshop 2020

TheProblemofFragmentedOcclusion inObjectDetection JulianPegoraro1,RomanPflugfelder1,2 1 AITAustrian InstituteofTechnology, 1,2 TUWien {julian.pegoraro|roman.pflugfelder}@ait.ac.at,roman.pflugfelder@tuwien.ac.at Abstract. Object detection in natural environments is still a very challenging task, even though deep learning has brought a tremendous improvement in performance over the last years. A fundamental problemofobjectdetectionbasedondeeplearningis thatneither the trainingdatanor thesuggestedmod- els are intended for the challenge of fragmented oc- clusion. Fragmented occlusion is muchmore chal- lenging than ordinary partial occlusion and occurs frequently innatural environments suchas forests. A motivating example of fragmented occlusion is ob- ject detection through foliage which is an essential requirement ingreenbordersurveillance. Thispaper presentsananalysisofstate-of-the-artdetectorswith imageryofgreenbordersandproposes to trainMask R-CNNonnew training datawhich captures explic- itly theproblemof fragmentedocclusion. Theresults show clear improvements ofMask R-CNNwith this new training strategy (also against other detectors) fordata showingslight fragmentedocclusion. 1. Introduction Automated surveillance at green borders has be- come a hot topic for European border guards. Bor- der guards today face several challenges in protect- ing EU borders. One well known occasion in public is illegalmigrationwhichhad itspeak in2015. Border surveillance today limited to 2D imag- ing sensors consists of color and thermal cameras, mounted on poles or used as handheld cameras by the border guards. Innovating these technical sys- tems by adding further capabilities of automatic in- ference such as the automatic detection of persons, vehicles, animals and suspicious objects in general will need to applyobjectdetectors to such imagery. However, video of green borders especially at EU borders show significant differences to typical im- agery of video surveillance such as indoor video or Figure 1: The problem of fragmented occlusion in object detection. Top Left: no occlusion (levelL0). Top Right: slight occ. (L1). Bottom Left: moderate occ. (L2). BottomRight: heavyocc. (L3) occlusion. video taken in man-made outdoor scenes. For exam- ple, green borders are scenes showing dense forest, hills, harsh weather and climate conditions. Such scenes draw challenges to automated surveillance and raise several interesting researchquestions. This paper considers a challenge for state-of- the-art object detection in green border surveillance which is the problem of through foliage detection. To the best of our knowledge, none of the current approaches for object detection allow the detection of objects through foliage. This problem raises an interesting scientific question, namely how to detect objects with fragmented occlusion? This problem is also different to the problem of partial occlusion in 98