MPLR-CapsNet: A novel capsule network with multi-line parallel features and logical reasoning for occluded pedestrian detection

Abstract

Occluded pedestrian detection is a key technology in computer vision. The parts of pedestrians covered are random, and diverse clothes, postures, and varying scales make occluded pedestrian detection a challenging task. In order to solve the above issues, we propose a novel capsule network model, MPLR-CapsNet, based on multi-line parallel features and logical reasoning, aiming to achieve accurate detection of occluded pedestrians in dense scenes. This network model introduces RS module to learn diversified key feature information in a directional manner, and use Mix-squash function to enhance the anti-interference and generalization abilities of the network model. Furthermore, the local pruning strategy is adopted to retain only the important low-level capsules connected to high-level capsules in the dynamic routing stage, effectively reducing the training parameters of the network model. In addition, we propose a capsule reasoning module based on the hyper-sausage metric model in the reasoning detection stage, transforming the object category and coordinates into the attribute representations of capsule entities, which significantly enhance the autonomous reasoning and discriminant abilities of the network model. Extensive experiments conducted on two standard occluded pedestrian datasets validate the advancement of the proposed method, outperforming existing state-of-the-art solutions. Our model achieves Top-1 accuracies of 97.6 % and 94.5 % on the CityPersons and CrowdHuman datasets, respectively.