Mobile Robot Tracking with Deep Learning Models under the Specific Environments

Abstract

Visual-based target tracking is one of the critical methodologies for the control problem of multi-robot systems. In dynamic mobile environments, it is common to lose the tracking targets due to partial visual occlusion. Technologies based on deep learning (DL) provide a natural solution to this problem. DL-based methods require less human intervention and fine-tuning. The framework has flexibility to be retrained with customized data sets. It can handle massive amounts of available video data in the target tracking system. This paper discusses the challenges of robot tracking under partial occlusion and compares the system performance of recent DL models used for tracking, namely you-only-look-once (YOLO-v5), Faster region proposal network (R-CNN) and single shot multibox detector (SSD). A series of experiments are committed to helping solve specific industrial problems. Four data sets are that cover various occlusion statuses are generated. Performance metrics of F1 score, precision, recall, and training time are analyzed under different application scenarios and parameter settings. Based on the metrics mentioned above, a comparative metric P is devised to further compare the overall performance of the three DL models. The SSD model obtained the highest P score, which was 13.34 times that of the Faster RCNN model and was 3.39 times that of the YOLOv5 model with the designed testing data set 1. The SSD model obtained the highest P scores, which was 11.77 times that of the Faster RCNN model and was 2.43 times that of the YOLOv5 model with the designed testing data set 2. The analysis reveals different characteristics of the three DL models. Recommendations are made to help future researchers to select the most suitable DL model and apply it properly in a system design.