• urban mobility
• traffic analysis
• object tracking
• object detection
• Computer vision

Computer Vision has become a pivotal component in Intelligent Transportation Systems (ITS) and traffic surveillance, responding to the demands of increasingly congested and densely populated urban environments. Harnessing the availability and ever-decreasing costs of video surveillance infrastructure and hardware applied to build and run Deep Neural Networks, more and more government bodies are looking into adopting these technologies. All the while keeping an eye on privacy and ethic concerns, which are inherently relevant to the topic at hand, especially nowadays within the General Data Protection Regulation (GDPR) implemented in Europe. To achieve compliance, secure paradigms such as edge computing and the proper handling of the data becomes an integral part of the solution.

This thesis presents a real-time traffic monitoring solution that encompasses vehicle and pedestrian detection, tracking and analysis of their interactions, then generating further insights like object count, speed and trajectory estimation, crowd control, collision detection and congestion identification. This is done by leveraging a custom YOLO Deep Convolutional Neural Network for object detection together with an enhanced tracking algorithm and a pose detection model which are simultaneously applied to a video stream in order to understand the events happening in real-time, extracting vital information about each frame such as the type and position of each object of relevance within the scene, also tracking their positions and building a history of the movement associated to each individual object, assigning them a unique ID. Below the deck, the solution is supported by Microsoft Azure’s architecture to handle every complementary step, such as the network necessary to retrieve the video streams, the training and deployment of machine learning models, the storage and control of the resulting metadata from the vision algorithms and the alerting system for risk events such as collisions or congestions.