In core computer vision tasks, we have witnessed significant advances in object detection, localisation and tracking. However, there are currently no methods to detect, localize and track objects in road environments, and taking into account real-time constraints. In this paper, our objective is to develop a deep learning multi object detection and tracking technique applied to road smart mobility. Firstly, we propose an effective detector-based on YOLOv3 which we adapt to our context. Subsequently, to localize successfully the detected objects, we put forward an adaptive method aiming to extract 3D information, i.e., depth maps. To do so, a comparative study is carried out taking into account two approaches: Monodepth2 for monocular vision and MADNEt for stereoscopic vision. These approaches are then evaluated over datasets containing depth information in order to discern the best solution that performs better in real-time conditions. Object tracking is necessary in order to mitigate the risks of collisions. Unlike traditional tracking approaches which require target initialization beforehand, our approach consists of using information from object detection and distance estimation to initialize targets and to track them later. Expressly, we propose here to improve SORT approach for 3D object tracking. We introduce an extended Kalman filter to better estimate the position of objects. Extensive experiments carried out on KITTI dataset prove that our proposal outperforms state-of-the-art approches.

In this article, we deal with the problem of understanding human-to-human interactions as a fundamental component of social events analysis. Inspired by the recent success of multi-modal visual data in many recognition tasks, we propose a novel approach to model dyadic interaction by means of features extracted from synchronized 3D skeleton coordinates, depth, and Red Green Blue (RGB) sequences. From skeleton data, we extract new view-invariant proxemic features, named Unified Proxemic Descriptor (UProD), which is able to incorporate intrinsic and extrinsic distances between two interacting subjects. A novel key frame selection method is introduced to identify salient instants of the interaction sequence based on the joints’ energy. From Red Green Blue Depth (RGBD) videos, more holistic CNN features are extracted by applying an adaptive pre-trained Convolutional Neural Networks (CNNs) on optical flow frames. For better understanding the dynamics of interactions, we expand the boundaries of dyadic interactions analysis by proposing a fundamentally new modeling for non-treated problem aiming to discern the active from the passive interactor. Extensive experiments have been carried out on four multi-modal and multi-view interactions datasets. The experimental results demonstrate the superiority of our proposed techniques against the state-of-the-art approaches.

In this paper, we introduce a novel and synergistic approach that combines attention mechanisms, LVENet for image visibility enhancement, and a tailored head pruning method for Multi-Head Self Attention (MHSA) models, specifically engineered for AACN (Attention Augmented Convolutional Network) and BoTNet (Bottleneck Transformers). The integration of these techniques aims to comprehensively address the challenges associated with object detection in the maritime domain. The attention mechanism selectively emphasizes critical areas of the image, Low-Visibility Enhancement Network (LVENet) enhances visibility under challenging conditions, and the head pruning method optimizes model efficiency and simplicity. Employing meticulous selection and evaluation, our approach achieves precise head pruning without compromising detection performance. Validation using common and maritime datasets underscores the effectiveness of our approach. The results showcase a substantial reduction in epoch time by over 30%, while enhancing accuracy, improving computational efficiency, and streamlining model complexity. This innovation facilitates deployment in challenging maritime scenarios.