Visual Object Detection
Object detection is one of the most fundamental and difficult problems in computer vision. It aims to discover object instances in real images from a huge number of established categories. Handcrafted characteristics were initially used to create the majority of the early object detection algorithms. People had no choice but to build complicated feature representations and a range of speed up skills to exhaust the use of limited computing resources due to the lack of effective image representation at the time. The last decade, deep learning has emerged as a powerful tool for learning feature representations directly from data, resulting in significant advances in the field of generic object detection. Convolutional Neural Networks (CNNs) are the most common deep learning models employed in the field of object detection right now. A typical CNN has a hierarchical structure and is made up of several layers that are used to learn data representations with multiple levels of abstraction. The number of convolutional layers and kernels in the CNN model can be increased to improve accuracy. However, increasing the number of convolutional layers demands more computational resources, otherwise it slows down the CNN model’s processing speed. This serious problem is considered in cases, where the object detection needs to be applied locally on embedded devices with lower computational resources in inaccessible areas. The low-power image recognition challenge emphasizes a balance of accuracy, throughput, and power budget. In certain sectors, such as the internet of things (IoT), robots, autonomous driving, and drone-based surveillance, these goals are not just appealing but also necessary.
In the context of the ISOLA project, a number of monitoring-related services will be exploited, which will handle multiple visual data streams from security cameras and visual sensors running as UxV payloads. CERTH will develop an object detection solution based on deep neural networks that will give state of the art accuracy and performance, while taking into account the available computational resources. There are cases, where the object detector will be used for monitoring within the ship and it will have available the resources of powerful and high-end hardware. However, there are additional scenarios, in which this object detector must be used at great distances from the ship, on an embedded device mounted on a UAV. CERTH’s algorithm will be able to return results with increased performance and accuracy, despite the limited processing resources now available. In this way, CERTH will ensure that the ship’s security officers will receive the best situational awareness from all the visual sources available at the minimum time.