Monocular Vision-based Real-Time Pallet Detection in Harsh Industrial Environment

It is the important global trend to use the unmanned production lines in order to meet the demand of customers and improve the efficiency for industrial processes. The automated storage and delivery system (ASDS) is one of the main components of the unmanned production line. It consists of many shields and several automata and complex control systems for loading and unloading, so its cost is so high. For the tradeoff of the cost and performance of cargo handling, forklift is a best alternative to the lack of financial ability enterprises/factories. In this paper, we propose a pallet detection method to allow forklifts to engage the pallet autonomously using only a monocular vision on the forklift in the harsh industrial environment. To reduce the number of features and increases the detection efficiency, we describe the pallet features by combining the Haar-like features and multi-block local binary pattern (MBLBP). 8 sets of Haar-type encoding models make the LBP feature better to encode the local structure. Adaboost classifier that use distribution information of features in training set, allows to detect pallet candidates with high accuracy and efficiency in harsh industrial environments. In particular, improved feature to maximize the margin when pattern classes are projected onto the classification hyperplane is used to enhance the discriminate ability of classifier and reduce the computational cost. The analysis of the geometric features of the pallets using integral-sum-difference (ISD) excludes the wrong candidates with high efficiency. The experimental results demonstrate that our proposed algorithm could detect the pallet with average rate of more than 98% and is robust to environmental changes.