The proliferation in surveillance cameras can be leveraged to alleviate crime by deploying an automated weapon (knife) detection system. This work presents a novel object detection algorithm and its application to visual knife detection for video data. The approach is tolerant to rotation, and change in scale and pose. Knife detection is a challenging problem mainly because of extensive variations in the shape, texture and size of knives. The proposed approach has three stages, foreground segmentation, Features from Accelerated Segment Test (FAST) based prominent feature detection for image localization and Multi-Resolution Analysis (MRA) for classification and target confirmation. The approach is scalable due to its client-server architecture, and achieves parallelism by doing the bulk of computation in the cloud. Empirical evaluation of the technique shows promising results as compared with the other approaches. Our contribution is the study of knife detection problem and development of a robust object detection algorithm.