In this thesis, we present a video object counting approach using multiple local feature matching. We explain the development of a dataset with which to test our approach. Our dataset uses a new approach which we designed to extract object ground truth. We also provide a comparison of common single object trackers. We develop a multi-object tracker named Learn-Select-Track and use it to track the colours of objects of interest to filter out false positive object localisations. We discuss the implementation of the HDBSCAN algorithm which we use in our novel approach for matching multiple local feature descriptors. We show that the detected clusters provide very good matches for the features and demonstrate our approach to cluster analysis and validation. We develop a simple yet efficient way of learning the features of the object of interest which is independent of the number of objects in the frame. We also develop a computationally simple way of detecting the other objects in the frame by using a combination of the detected clusters, the features of the object of interest and vector algebra. Our approach is capable of detecting partially visible and occluded objects as well. We present three ways of extracting object count estimations from videos and provide empirical evidence to show that our approach can be used in a wide variety of scenarios.