Intelligent environment monitoring confronts the major challenge of electrical power minimization with the constraints of real-time response and high quality of service. Such systems are mostly battery-operated and typically perform the tasks of monitoring various (complex) physical phenomena (e.g., forest fire, solar radiation, crowed monitoring, intelligent traffic surveillance, border security, etc) in a hazardous environment and must have adequate processing capabilities to meet the real-time response requirements. With these capabilities being the focus of design, high power consumption is inevitable which is not affordable in the intended environments of operation. A system must be designed with power-efficient components as well as online power optimization strategies. This paper presents a critical analysis of the requirements and constraints for power and resource awareness in environment monitoring systems. We classify the existing power management techniques used in vision-based environment monitoring systems and identify the set of features that serves as a criterion to evaluate the performance of a technique. Our critical analysis and experiments with carefully designed systems demonstrate that a machine-learning based power optimization strategy gives optimal performance in dynamic environments and outperforms other techniques.