Vehicles detection and tracking using Convolutional Neural Networks on edge-computing platforms

Image classification and detection tasks have been approached in numerous ways throughout the years. Traditional approaches used conventional machine learning techniques dealing with hand-crafted vectors of features which had to be carefully extracted from the images with considerable domain-expertise. Deep learning algorithms have overcome this limitation by providing architectures able to work on raw data while outperforming traditional techniques. A deep learning architecture which proved to be very successful in the context of computer vision is the Convolutional Neural Network, an architecture able to achieve previously unreachable performances, making it, nowadays, a de-facto standard in classification and detection tasks. However, a drawback of deep learning architectures is their high computational footprint which has limited, until recently, their adoption only to high-performance servers and workstations. The availability of fast and efficient CNN models together with the release of AI accelerating low-powered platforms has enabled the use of CNNs on the edge, empowering a shift from a server-centric scenario to a fog scenario, lowering data bandwidth requirements for connected nodes and enabling smart features on any connected device. This work addresses the design and implementation of a CNN based framework for performing road traffic analysis through vehicles detection, classification and tracking, able to efficiently run on a low-powered edge-computing platform. The resulting platform can be positioned in close proximity to the node, enabling these smart-features on any traditional IP camera. The first efforts of this work went into selecting the most promising CNN models and fine-tune them on a vehicles specific dataset, evaluating the results both in terms of speed and detection performance. With a strong and efficient model, the later efforts went into developing the framework components: a detection pipeline, a REST server and a React web app. The design embraced a flow-based paradigm, keeping a constant eye on modularity and extensibility, making it easily tailorable to completely different scenarios.