Optimising underlying processes and increasing work safety for employees are two key concerns for the warehouse industry. Common technologies for warehouses are warehouse management systems (WMS) and radio frequency identifications (RFID). WMS aims at handling warehouse operations in an improved way and to store relevant inventory information. RFID enables products to be automatically identified, resulting in cost savings through shorter handling timelines. Despite the benefits associated with WMS or RFID, there is still a margin for improvement in the commonly employed technologies used in the warehouse industry.

Computer vision algorithms have been used to solve complex video-based analytical solutions, but their potential in warehouse surveillance applications is not yet fully explored. The recent use of deep learning-based architectures has brought a revolution in the global information technology industry. When supplied with a bulk of training data and with the availability of fast computational resources like graphical processing units (GPUs), these models have brought wonders in improving detection and localisation accuracy of objects of interest. These models have successfully been used in solving a wide diversity of computer vision tasks.

This project aims at exploring and studying the multiple object tracking (MOT) problem in the context of a typical warehouse environment. Industrial vehicles (i.e. forklift trucks) and pedestrians are detected and tracked in a complex warehouse environment featuring challenges like occlusions, clutter or illuminations variations. In industrial warehouses, a robust multiple object tracking framework could provide useful information on a forklift truck and pedestrian movements, to be used for processes improvement. It can help in enhancing situational awareness in warehouses and increase safety levels of the pedestrians working there.

This thesis describes the implementation, evaluation and analysis designed and selected for model-based computer vision algorithms for tracking multiple targets. Owing to the success of deep feature-based tracking mechanisms, we train, evaluate and explore their potential in a warehouse environment using a detection based tracking approach. Detection to track association methods are revisited, with a novel proposed track association algorithm. The use of various hand-crafted and deep feature-based methods and their limitations in different tracking mechanisms are studied in detail. A novel composite feature-based ensemble tracker is also proposed, leading to robust visual tracking results.

This thesis also reports about an experimental comparison for robust multiple object trackers for warehouse environments, discussing model-based versus deep learning-based methods, thus highlighting their limitations and future research directions.