Detection of Package Edges in Distance Maps

Abstract

This paper presents a CNN-based algorithm for detecting package edges in a scene represented with a distance map (range image), trained on a custom dataset of packaging scenarios. The proposed algorithm represents the basis for package recognition for automatic trailer loading/unloading. The main focus of this paper is designing a semantic segmentation CNN model capable of detecting different types of package edges in a distance map containing distance errors characteristic of Time-of-Flight (ToF) scanning, and differentiating box edges from edges belonging to other types of packaging objects (bags, irregular objects, etc.). The proposed CNN is optimized for training with a limited number of samples containing heavily imbalanced classes. Generating a binary mask of edges with 1-pixel thickness from the probability maps outputted from the CNN is achieved through a custom non-maximum suppression-based edge thinning algorithm. The proposed algorithm shows promising results in detecting box edges.