About one-third of adults in the U.S. aged 65 or older fall every year, with 20% of the reported fall cases needing prompt medical attention. The methods that have been proposed for fall detection in recent years present trade-offs between level of intrusiveness, coverage area, processing power requirements and detection accuracy. We present a robust and resource-efficient method for fall detection by using a wearable embedded smart camera, which is a small, battery-operated unit. The proposed approach employs histograms of edge orientations as well as edge strength values, and analyzes their correlation. Moreover, we adaptively determine the cells that do not contribute to overall edge information, and remove them autonomously. Since the camera is worn by the subject, monitoring can continue wherever the subject may go including outdoors. The captured frames are not the images of the subject, and this alleviates the privacy concerns. The alert and an image of the surroundings can be transmitted wirelessly, only when a fall event is detected, for easier localization of the subject by emergency response teams. The experimental results obtained with over 300 trials are very promising with a 91% detection rate for falls.