This thesis looks to advance understanding in the field of computer vision based crowd analysis through a combination of deep learning techniques, multi-task learning, and domain adaptation. Issues that have limited progress in this field to date include visual occlusion, scale and perspective issues, variation in scene content as well as a lack of labelled training data. Another negative trend that has emerged in this field as well as in computer vision in general is the development of bespoke, single-task techniques that cannot be easily extended or re-used. The core contributions of this work are as follows. First, deep learning methods are developed for several crowd analysis tasks including crowd counting, crowd density level estimation, crowd behaviour recognition and crowd behaviour anomaly detection. The proposed data-driven methods are shown to be superior to techniques which rely on hand-crafted features, overcoming many of the observed challenges and achieving state-of-the-art results. Second, multi-task learning strategies are applied to crowd behaviour and congestion analysis tasks, increasing the overall predictive performance and removing redundant model parameters. Finally, domain adaptation techniques are investigated as a means to extend a given crowd analysis model to perform the same task in new visual domains (e.g. medical, wildlife) and vice-versa, with original domain performance preserved.