Nerea Aranjuelo

Bio: Nerea Aranjuelo is an academic researcher. The author has contributed to research in topics: Computer science & Deep learning. The author has an hindex of 1, co-authored 1 publications receiving 77 citations.

Multi-camera BEV video-surveillance system for efficient monitoring of social distancing

Abstract: The current sanitary emergency situation caused by COVID-19 has increased the interest in controlling the flow of people in indoor infrastructures, to ensure compliance with the established security measures. Top view camera-based solutions have proven to be an effective and non-invasive approach to accomplish this task. Nevertheless, current solutions suffer from scalability problems: they cover limited range areas to avoid dealing with occlusions and only work with single camera scenarios. To overcome these problems, we present an efficient and scalable people flow monitoring system that relies on three main pillars: an optimized top view human detection neural network based on YOLO-V4, capable of working with data from cameras at different heights; a multi-camera 3D detection projection and fusion procedure, which uses the camera calibration parameters for an accurate real-world positioning; and a tracking algorithm which jointly processes the 3D detections coming from all the cameras, allowing the traceability of individuals across the entire infrastructure. The conducted experiments show that the proposed system generates robust performance indicators and that it is suitable for real-time applications to control sanitary measures in large infrastructures. Furthermore, the proposed projection approach achieves an average positioning error below 0.2 meters, with an improvement of more than 4 times compared to other methods.

Leveraging Synthetic Data for DNN-Based Visual Analysis of Passenger Seats

Abstract: Deep neural network (DNN)-based vision systems could improve passenger transportation safety by automating processes such as verifying the correct positioning of luggage, seat occupancy, etc. Abundant and well-distributed data are essential to make DNNs learn appropriate pattern recognition features and have enough generalization ability. The use of synthetic data can reduce the effort of generating varied and annotated data. However, synthetic data usually present a domain gap with real-world samples, that can be reduced with domain adaptation techniques. This paper proposes a methodology to build simulated environments to generate balanced and varied synthetic data and avoid including redundant samples to train classification DNNs for passenger seat analysis. We show a practical implementation for detecting whether luggage is correctly placed or not in an aircraft cabin. Experimental results show the contribution of the synthetic samples and the importance of correctly discarding redundant data.

Methodology for generating synthetic labeled datasets for visual container inspection

Abstract: Nowadays, containerized freight transport is one of the most important transportation systems that is undergoing an automation process due to the Deep Learning success. However, it suffers from a lack of annotated data in order to incorporate state-of-the-art neural network models to its systems. In this paper we present an innovative methodology to generate a realistic, varied, balanced, and labelled dataset for visual inspection task of containers in a dock environment. In addition, we validate this methodology with multiple visual tasks recurrently found in the state of the art. We prove that the generated synthetic labelled dataset allows to train a deep neural network that can be used in a real world scenario. On the other side, using this methodology we provide the first open synthetic labelled dataset called SeaFront available in: https://datasets.vicomtech.org/di21-seafront/readme.txt.

Normalization in Training U-Net for 2-D Biomedical Semantic Segmentation

Abstract: Two-dimensional (2-D) biomedical semantic segmentation is important for robotic vision in surgery. Segmentation methods based on deep convolutional neural network (DCNN) can out-perform conventional methods in terms of both accuracy and levels of automation. One common issue in training a DCNN for biomedical semantic segmentation is the internal covariate shift where the training of convolutional kernels is encumbered by the distribution change of input features, hence both the training speed and performance are decreased. Batch normalization (BN) is the first proposed method for addressing internal covariate shift and is widely used. Instance normalization (IN) and layer normalization (LN) have also been proposed. Group normalization (GN) is proposed more recently and has not yet been applied to 2-D biomedical semantic segmentation (GN was used in 3-D biomedical semantic segmentation in [P.-Y. Kao, T. Ngo, A. Zhang, J. Chen, and B. Manjunath, Brain tumor segmentation and tractographic feature extraction from structural MR images for overall survival prediction 2018, arXiv:1807.07716], however, no specific validations on GN were given). Most DCNNs for biomedical semantic segmentation adopt BN as the normalization method by default, without reviewing its performance. In this letter, four normalization methods—BN, IN, LN, and GN are compared in details, specifically for 2-D biomedical semantic segmentation. U-Net is adopted as the basic DCNN structure. Three datasets regarding the right ventricle, aorta, and left ventricle are used for the validation. The results show that detailed subdivision of the feature map, i.e., GN with a large group number or IN, achieves higher accuracy. This accuracy improvement mainly comes from better model generalization. Codes are uploaded and maintained at Xiao-Yun Zhou's Github.