CNN-based track detection in coherent SAR imagery: a distinction between foot and vehicle tracks

The goal of this paper is to assess the capability of a CNN based algorithm to detect vehicle and foot tracks and to distinguish them. The used CNN architecture has already proven to be very effective for the segmentation of vehicle tracks in previous work and thus was chosen for this investigation. Foot tracks in general are of a poorer distinctness and not as linear or constant as vehicle tracks. Also, when heavily overlapping, the signatures of foot tracks loose some of their structural features and instead are more likely to be distinguishable by their texture. Thus, two approaches for segmentation labeling are investigated: First, a line-based labeling, which considers the individual tracks; and second a region-based labeling, allowing for textural features. How well these two labeling approaches perform, is tested and results are shown regarding the detection of foot tracks and their distinction from vehicle tracks.