The RECODists Karina Olga Maizman Bogdan and Laurindo de Sousa Britto Neto have earned their master’s and Ph.D. degrees, respectively. Congratulations!
Karina presented the dissertation entitled “Multiple Tracklet Matching under Severe Occlusions” supervised by Prof. Siome Goldenstein. Laurindo presented the thesis entitled “Wearable Systems based on Computer Vision Methods to aid the Visually Impaired People” also supervised by Prof. Siome Goldenstein and co-supervised by Profa. Maria Cecília Baranauskas. Both abstracts are shown below.
Multiple Tracklet Matching under Severe Occlusions
Tracking is an important area of computer vision, and despite the recent progress by the several proposed trackers many challenges remain unsolved. One of them is occlusion which affects the correctness and continuity of the object’s trajectory, from single to multiple object environments. In the latter, the difficulty level of tracking is increased since the method must also be capable of maintaining the objects identities. As most of the current methods rely on detection responses, an event of occlusion generally increase the detection missing rate, fragmenting the trajectories into tracklets in data association approaches that perform subsequent association over these responses. Particular classes of the event such as full and long-term occlusions involving several targets make the problem even more challenging. Furthermore, in complex scenarios such as the sport-based context tracking multiple objects involves dealing with similar appearances and motion, besides the high-dynamic natural trait of the environment. To address this problem we propose a tracklet matching method formulated as a graph-based model. We evaluate different representations for the three main components of our model: the tracklet representation related to the trajectory and the appearance model of the corresponding target, the association cost that define the affinity over two tracklets, and the matching algorithm itself. Different from current literature, we evaluate our proposed model in a new tracking dataset with a multi-camera setup that includes severe occlusion cases in a sport-based context.
Wearable Systems based on Computer Vision Methods to aid the Visually Impaired People
Blind or low vision people suffer many difficulties in performing daily activities. Efforts have been expended for the design of assistive systems that facilitate the attainment of specific tasks for the visually impaired people. In particular, the field of Computer Vision has a lot to contribute. In this Ph.D. Thesis, we proposed a real-time face recognition approach, simple and efficient for the development of wearable systems with limited hardware resources, within the scope of applications that are designed to aid blind and low vision people. The proposed approach has been tested and validated through the development of two systems, which acted as case studies of this research. Such systems used as main hardware components of the sensor Microsoft Kinect for Windows (Kinect) and the smartwatch Samsung Galaxy Gear (Gear). The Kinect-based system uses the sensor as a wearable device, performs face detection and uses temporal coherence along with a simple biometric procedure to generate a sound associated with the identified person, virtualized at his/her estimated 3D location. Likewise, the Gear-based system performs all people identification procedure, with minor differences in technology, but without estimating the 3D location. Performance/Accuracy Experiments were performed and compared with both traditional and state-of-the-art face recognition approaches. The validation used a new dataset of RGB-D video data, collected with the Kinect, called Unicamp Kinect Face Database, with 600 videos of 30 people, containing variation of illumination, background and movement patterns, and that will be made publicly-available. Experiments with others existing datasets in the literature were also considered. Battery consumption was estimated through experiments using the hardware of the smartwatch. The results show that our approach, on average, outperforms traditional face recognition methods while requiring much less computational resources (memory, processing power and battery life) when compared to existing techniques in the literature, deeming it suitable for the wearable hardware constraints and the limited-scope applications. In addition, an alternative people recognition approach in the dark, based in original approach, has been proposed using depth-only data from Kinect. Finally, experiments were conducted to evaluate the user’s interaction with the systems both with blind and low vision users as with blindfolded users to find critical problems, causing improvements in systems and verifying the best ways to perform the audio feedback. Such experiments have shown encouraging results.