Computer-vision (CV) applications are an important part of mobile robot automation, analyzing the perceived raw data from vision sensors and providing a rich amount of information on the surrounding environment. The design of a high-speed and energy-efficient CV application for a resource-constrained mobile robot, while maintaining a certain targeted level of accuracy in computation, is a challenging task. This is because such applications demand a lot of resources, e.g. computing capacity and battery energy, to run seamlessly in real time. Moreover, there is always a trade-off between accuracy, performance and energy consumption, as these factors dynamically affect each other at runtime. In this thesis, we investigate novel runtime resource management approaches to improve performance and energy efficiency of vision-based applications in mobile robots. Due to the dynamic correlation between different management objectives, such as energy consumption and execution time, both environmental and computational observations need to be dynamically updated, and the actuators are manipulated at runtime based on these observations. Algorithmic and computational parameters of a CV application (output accuracy and CPU voltage/frequency) are adjusted by measuring the key factors associated with the intensity of computations and strain on CPUs (environmental complexity and instantaneous power). Furthermore, we show how mechanical characteristics of the robot, i.e. the speed of movement in this thesis, can affect the computational behaviour. Based on this investigation, we add the speed of a robot, as an actuator, to our resource management algorithm besides the considered computational knobs (output accuracy and CPU voltage/frequency). To evaluate the proposed approach, we perform several experiments on an unmanned ground vehicle equipped with an embedded computer board and use RGB and event cameras as the vision sensors for CV applications. The obtained results show that the presented management strategy improves the performance and accuracy of vision-based applications while significantly reducing the energy consumption compared with the state-of-the-art solutions. Moreover, we demonstrate that considering simultaneously both computational and mechanical aspects in management of CV applications running on mobile robots significantly reduces the energy consumption compared with similar methods that consider these two aspects separately, oblivious to each other’s outcome.