Multi-Object Tracking (MOT) techniques have been under continuous research and increasingly applied in a diverse range of tasks. One area in particular concerns its application in navigation tasks of assistive mobile robots, with the aim to increase the mobility and autonomy of people suffering from mobility decay, or severe motor impairments, due to muscular, neurological, or osteoarticular decay. Therefore, in this work, having in view navigation tasks for assistive mobile robots, an evaluation study of two MOTs by detection algorithms, SORT and Deep-SORT, is presented. To improve the data association of both methods, which are solved as a linear assignment problem with a generated cost matrix, a set of new object tracking data association cost matrices based on intersection over union, Euclidean distances, and bounding box metrics is proposed. For the evaluation of the MOT by detection in a real-time pipeline, the YOLOv3 is used to detect and classify the objects available on images. In addition, to perform the proposed evaluation aiming at assistive platforms, the ISR Tracking dataset, which represents the object conditions under which real robotic platforms may navigate, is presented. Experimental evaluations were also carried out on the MOT17 dataset. Promising results were achieved by the proposed object tracking data association cost matrices, showing an improvement in the majority of the MOT evaluation metrics compared to the default data association cost matrix. In addition, promising frame rate values were attained by the pipeline composed of the detector and the tracking module.

This paper focuses on the task of RGB indoor scene classification. A single scene may contain various configurations and points of view, but there are a small number of objects that can characterize the scene. In this paper we propose a deep-learning based Global and Semantic Feature Fusion Approach (GSF ² App) with two branches. In the first branch (top branch), a CNN model is trained to extract global features from RGB images, taking leverage from the ImageNet pre-trained model to initialize our CNN’s weights. In the second branch (bottom branch), we develop a semantic feature vector that represents the objects in the image, which are detected and classified through the COCO dataset pre-trained YOLOv3 model. Then, both global and semantic features are combined in an intermediate feature fusion stage. The proposed approach was evaluated on the SUN RGB-D Dataset and NYU Depth Dataset V2 achieving state-of-the-art results on both datasets.

Robot localization in indoor environments is crucial for achieving flexible automated navigation. In this paper, we propose a novel multi-stage localization approach for mobile robots which combines a commercial beacon-based absolute indoor positioning system with laser scan data. This configuration of sensors can be deployed in a non-intrusive way in most robotic platforms without the use of proprioceptive sensors (e.g. wheel encoders) which often introduce non-negligible maintenance and downtime costs. The data fusion is performed by a particle filter aided by a refinement stage. The proposed approach was evaluated in two different indoor scenarios with a mobile platform equipped with a mobile Marvelmind beacon and a Hokuyo UTM-30LX scanning laser rangefinder. The proposed localization framework, purposely without proprioceptive data, is compared with an AMCL approach having as inputs odometry, calculated from wheel encoders' data, and laser scan data. Preliminary results show that the proposed approach can provide an accurate localization estimate and that using the refinement stage improves localization.

This paper introduces a new Reinforcement Learning (RL) based local motion planning (RL-LMP) approach for mobile robots. This method enables virtual or real mobile platforms, to follow a path and navigate from location A to B. The method consists of two stages: a training stage and an online stage. The training stage consists in the robot iteratively learning to follow previously defined paths in a simulation environment. New RL state representations are proposed as well as a strategy to deal with the delayed reward problem and an approach to guide the exploration-exploitation of the RL model considering prior knowledge. Through training in a virtual environment and assimilation of human driving behaviors (using a gamepad), an RL model is obtained and used in the online stage, enabling a mobile platform, in a simulation or real environment, to move along a path avoiding obstacles. A set of tests and experiments were performed in different scenarios in virtual and real environments in order to assess the performance of the local motion planning approach. Validation of the RL-LMP approach, in a real environment, was carried out in the ISR-InterBot platform with the algorithm developed in ROS. The obtained results show that the RL-LMP approach provides a valid solution to the local motion planning problem (small lateral error while following paths) and suggests promising perspectives for improvement in the future.