Development of obstacle detection and tracking for motion-based verification crucially depends on datasets that reflect real-world challenges and on the performance measures that have strong algorithm probing properties. In the referenced paper we have proposed a new long-term tracking dataset as well as new performance measures. The new measures outperform existing ones in interpretation potential and in better distinguishing between different tracking behaviors, while the dataset is the most challenging long-term tracking dataset with many target disappearances. The performance measures have become part of the standard evaluation protocol in the largest visual object tracking challenge in computer vision, VOT. In addition, several more datasets have been contributed to the VOT. Apart from the tracking datasets, we have also developed obstacle detection datasets and protocols [4,6,17] for performance evaluation, as well as for training deep obstacle detection networks [7].
COBISS.SI-ID: 1538564803
The paper describes the method of tracking waterborne obstacles using primarily 3D point cloud data obtained from the onboard stereo camera system. It combines segmentation of the point cloud to water surface and outliers (obstacle candidates) and the tracking of obstacles. Estimation of water surface is done independently from other sensors, as long as the method converges in predefined time; if it does not, the roll and pitch data from the inertial measurement unit are used to provide rough approximation. The segmentation is further aided by semantic segmentation of RGB images of the same scene, which eliminates image areas that are irrelevant for navigation and thus reduces the number of false positives. The tracking of segmented obstacles is facilitated by the novel method of “depth fingerprinting”, where tracks of objects are composed based on a metric that combines spatial relations between object detections and the detected object depth histogram as a means of identification of the same objects that appear through consecutive images. The end result is a waterborne obstacle detection method for unmanned surface vehicles with significantly reduced number of false positive detections. During the project work, it became apparent that stereo camera system onboard USV is a subject to frequent calibration due to mechanical stress caused by operation or transfer to or from the winter storage.This was further addressed by development of introspective, target-less self-calibration method, published [17].
COBISS.SI-ID: 12642388
Part of our research has focused on developing methods for image-based obstacle detection. In this paper we proposed a graphical model based on a Markov random field that incorporates the on-board inertial measurement unit (IMU) and a stereo verification algorithm that consolidates tentative detections obtained from the segmentation. The IMU constraints the prior probabilities on pixels and adjusts the spatial hiperpriors on the semantic components. In addition we developed a fast camera-IMU calibration algorithm and IMU-based horizon projection in the image. The method was later extended to stereo early fusion graphical model with significant performance gains and published at premiere robotics conference IROS2018 [6]. We have successfully transferred the new insights to deep networks. We proposed a new deep architecture that implements the concepts learned in this research and is currently the best performing USV obstacle detection method -- the results were published at a premiere robotics conference ICRA2020 [8]. While the latter work is preliminary, it forms a solid basis for our further research in deep-learning based obstacle detection methods.
COBISS.SI-ID: 1537746627
Part of our research focused on efficient and robust visual object trackers. In this paper we introduced the channel and spatial reliability concepts to discriminative correlation filter tracking and provided a novel learning algorithm for its efficient and seamless integration in the filter update and the tracking process. The method allows training from a large training range and increases the detection range. We have implemented a C++ version of the tracker, which does not require specialized hardware like GPUs and runs realtime on a single CPU. The tracker has achieved first place among fifty trackers on the VOT realtime challenge VOT-RT2017 and is currently the most robust CPU-based tracker available in the prominent computer vision library OpenCv. The real-time capability makes it particularly suitable for autonomous robots that require timely reaction for safe navigation in a dynamic open world. The tracker has been extended to long-term tracking and presented as an oral presentation at ACCV2018 [15] (top 4.5% submissions). The tracker was also extended to RGB+depth tracking by reconstruction and presented at premiere computer vision conference ICCV2019 [25].
COBISS.SI-ID: 1537691075