Utterance Level Feature Aggregation with Deep Metric Learning for Speech Emotion Recognition.

Emotion is a form of high-level paralinguistic information that is intrinsically conveyed by human speech. Automatic speech emotion recognition is an essential challenge for various applications; including mental disease diagnosis; audio surveillance; human behavior understanding; e-learning and human-machine/robot interaction. In this paper, we introduce a novel speech emotion recognition method, based on the Squeeze and Excitation ResNet (SE-ResNet) model and fed with spectrogram inputs. In order to overcome the limitations of the state-of-the-art techniques, which fail in providing a robust feature representation at the utterance level, the CNN architecture is extended with a trainable discriminative GhostVLAD clustering layer that aggregates the audio features into compact, single-utterance vector representation. In addition, an end-to-end neural embedding approach is introduced, based on an emotionally constrained triplet loss function. The loss function integrates the relations between the various emotional patterns and thus improves the latent space data representation. The proposed methodology achieves 83.35% and 64.92% global accuracy rates on the RAVDESS and CREMA-D publicly available datasets, respectively. When compared with the results provided by human observers, the gains in global accuracy scores are superior to 24%. Finally, the objective comparative evaluation with state-of-the-art techniques demonstrates accuracy gains of more than 3%.

DEEP-SEE: Joint Object Detection, Tracking and Recognition with Application to Visually Impaired Navigational Assistance.

In this paper, we introduce the so-called DEEP-SEE framework that jointly exploits computer vision algorithms and deep convolutional neural networks (CNNs) to detect, track and recognize in real time objects encountered during navigation in the outdoor environment. A first feature concerns an object detection technique designed to localize both static and dynamic objects without any a priori knowledge about their position, type or shape. The methodological core of the proposed approach relies on a novel object tracking method based on two convolutional neural networks trained offline. The key principle consists of alternating between tracking using motion information and predicting the object location in time based on visual similarity. The validation of the tracking technique is performed on standard benchmark VOT datasets, and shows that the proposed approach returns state-of-the-art results while minimizing the computational complexity. Then, the DEEP-SEE framework is integrated into a novel assistive device, designed to improve cognition of VI people and to increase their safety when navigating in crowded urban scenes. The validation of our assistive device is performed on a video dataset with 30 elements acquired with the help of VI users. The proposed system shows high accuracy (>90%) and robustness (>90%) scores regardless on the scene dynamics.

When Ultrasonic Sensors and Computer Vision Join Forces for Efficient Obstacle Detection and Recognition.

In the most recent report published by the World Health Organization concerning people with visual disabilities it is highlighted that by the year 2020, worldwide, the number of completely blind people will reach 75 million, while the number of visually impaired (VI) people will rise to 250 million. Within this context, the development of dedicated electronic travel aid (ETA) systems, able to increase the safe displacement of VI people in indoor/outdoor spaces, while providing additional cognition of the environment becomes of outmost importance. This paper introduces a novel wearable assistive device designed to facilitate the autonomous navigation of blind and VI people in highly dynamic urban scenes. The system exploits two independent sources of information: ultrasonic sensors and the video camera embedded in a regular smartphone. The underlying methodology exploits computer vision and machine learning techniques and makes it possible to identify accurately both static and highly dynamic objects existent in a scene, regardless on their location, size or shape. In addition, the proposed system is able to acquire information about the environment, semantically interpret it and alert users about possible dangerous situations through acoustic feedback. To determine the performance of the proposed methodology we have performed an extensive objective and subjective experimental evaluation with the help of 21 VI subjects from two blind associations. The users pointed out that our prototype is highly helpful in increasing the mobility, while being friendly and easy to learn.