Tracking multiple objects in very crowded scenes remains an extremely challenging task. This severely limits the degree of automation that can be achieved in a range of applications from pedestrian crowd management to ecosystem monitoring.  There is a rich literature on visual tracking, and many applications have been successfully deployed, including in self-driving cars. However, published methods tend to fail when scenes become very crowded and distinct individual-based visual features are simultaneously sparse. Recent advances in machine learning (deep learning) open up promising new alleys to tackle this problem. The project will explore how far we can push tracking in very crowded scenes using deep learning segmentation and classification combined with advanced optimisation methods.