Michel Meneses

Michel Meneses, Bruno Iwami

Modern approaches for keyword spotting rely on training deep neural networks on large static datasets with i.i.d. distributions. However, the resulting models tend to underperform when presented with changing data regimes in real-life applications. This work investigates a simple but effective online continual learning method that updates a keyword spotter on-device via SGD as new data becomes available. Contrary to previous research, this work focuses on learning the same KWS task, which covers most commercial applications. During experiments with dynamic audio streams in different scenarios, that method improves the performance of a pre-trained small-footprint model by 34%. Moreover, experiments demonstrate that, compared to a naive online learning implementation, conditional model updates based on its performance in a small hold-out set drawn from the training distribution mitigate catastrophic forgetting.

Michel Meneses, Leonardo Matos, Bruno Prado et al.

With the recent advances in the object detection research field, tracking-by-detection has become the leading paradigm adopted by multi-object tracking algorithms. By extracting different features from detected objects, those algorithms can estimate the objects' similarities and association patterns along successive frames. However, since similarity functions applied by tracking algorithms are handcrafted, it is difficult to employ them in new contexts. In this study, it is investigated the use of artificial neural networks to learning a similarity function that can be used among detections. During training, the networks were introduced to correct and incorrect association patterns, sampled from a pedestrian tracking data set. For such, different motion and appearance features combinations have been explored. Finally, a trained network has been inserted into a multiple-object tracking framework, which has been assessed on the MOT Challenge benchmark. Throughout the experiments, the proposed tracker matched the results obtained by state-of-the-art methods, it has run 58\% faster than a recent and similar method, used as baseline.

Michel C. Meneses, Leonardo N. Matos, Bruno O. Prado

This work presents a low-cost robot, controlled by a Raspberry Pi, whose navigation system is based on vision. The strategy used consisted of identifying obstacles via optical flow pattern recognition. Its estimation was done using the Lucas-Kanade algorithm, which can be executed by the Raspberry Pi without harming its performance. Finally, an SVM-based classifier was used to identify patterns of this signal associated with obstacles movement. The developed system was evaluated considering its execution over an optical flow pattern dataset extracted from a real navigation environment. In the end, it was verified that the acquisition cost of the system was inferior to that presented by most of the cited works, while its performance was similar to theirs.

Michel Conrado Cardoso Meneses

Autonomous robots are used as the tool to solve many kinds of problems, such as environmental mapping and monitoring. Either for adverse conditions related to the human presence or even for the need to reduce costs, it is certain that many efforts have been made to develop robots with an increasingly high level of autonomy. They must be capable of locomotion through dynamic environments, without human operators or assistant systems' help. It is noted, thus, that the form of perception and modeling of the environment becomes significantly relevant to navigation. Among the main sensing methods are those based on vision. Through this, it is possible to create highly-detailed models about the environment, since many characteristics can be measured, such as texture, color, and illumination. However, the most accurate vision-based navigation techniques are computationally expensive to run on low-cost mobile platforms. Therefore, the goal of this work was to develop a low-cost robot, controlled by a Raspberry Pi, whose navigation system is based on vision. For this purpose, the strategy used consisted in identifying obstacles via optical flow pattern recognition. Through this signal, it is possible to infer the relative displacement between the robot and other elements in the environment. Its estimation was done using the Lucas-Kanade algorithm, which can be executed by the Raspberry Pi without harming its performance. Finally, an SVM based classifier was used to identify patterns of this signal associated with obstacles movement. The developed system was evaluated considering its execution over an optical flow pattern dataset extracted from a real navigation environment. In the end, it was verified that the processing frequency of the system was superior to the others. Furthermore, its accuracy and acquisition cost were, respectively, higher and lower than most of the cited works.