Scale Invariant low frame rate tracking

The cameras used for road surveillance present low image quality and significant variability of objects' scale, making real-time vehicle detection and tracking challenging. This paper proposes Scale Invariant Low Frame Rate Tracking, called SILFT, a new methodology to track vehicles in most surveillance systems whose frame rate and resolution are low. This architecture uses the fusion of a dense optical flow with object detection in a 6-step pipeline. The detections are masks for the optical flow that later will be used for finding the object correspondence between frames, making it effective in low frame rate sequences with large object scale variability. SILFT achieved average precision for detection of 65.97%, surpassing the YOLO and FASTER R-CNN models in the proposed database. For the tracking task, SILFT achieved a PR-MOTA of 0.45, whereas the YOLO model with intersection over union tracking achieved 0.13 in the proposed dataset.