Saliency Detection with Moving Camera via Background Model Completion

Detecting saliency in videos is a fundamental step in many computer vision systems. Saliency is the significant target(s) in the video. The object of interest is further analyzed for high-level applications. The segregation of saliency and the background can be made if they exhibit different visual cues. Therefore, saliency detection is often formulated as background subtraction. However, saliency detection is challenging. For instance, dynamic background can result in false positive errors. In another scenario, camouflage will result in false negative errors. With moving cameras, the captured scenes are even more complicated to handle. We propose a new framework, called saliency detection via background model completion (SD-BMC), that comprises a background modeler and a deep learning background/foreground segmentation network. The background modeler generates an initial clean background image from a short image sequence. Based on the idea of video completion, a good background frame can be synthesized with the co-existence of changing background and moving objects. We adopt the background/foreground segmenter, which was pre-trained with a specific video dataset. It can also detect saliency in unseen videos. The background modeler can adjust the background image dynamically when the background/foreground segmenter output deteriorates during processing a long video. To the best of our knowledge, our framework is the first one to adopt video completion for background modeling and saliency detection in videos captured by moving cameras. The F-measure results, obtained from the pan-tilt-zoom (PTZ) videos, show that our proposed framework outperforms some deep learning-based background subtraction models by 11% or more. With more challenging videos, our framework also outperforms many high-ranking background subtraction methods by more than 3%.

Abstract P146: Assessment of Telestroke Sub-Events and Their Contribution to Door-To-Needle Time in a Telestroke Network

Introduction: Data remains scarce on which telestroke related sub-events (component parts/time intervals) are associated with delays in door-to-needle (DTN) time and goals for each telestroke sub-event. We aimed to assess the telestroke sub-events that contribute to DTN. After establishing set goals for each sub-event, we further evaluated the odds of DTN within 45 minutes if sub-event goals were achieved. Methods: We retrospectively analyzed prospectively collected data from a hub-and-spoke model telestroke network from January 2017 to September 2019. To determine which sub-events significantly contributed to DTN time, a sequential multiple regression analysis was performed. We entered covariates (age, sex, time of telestroke [day or night], NIHSS, average number of telestroke consults at a given site) in the first block followed by sub-events (door-to-telestroke request, door-to-CT, request-to-page, stroke physician response time, telestroke phone-to-video, video duration prior to needle and video completion-to-needle) in the second block. Logistic regression models were performed to estimate the odds of achieving a DTN within 45 minutes if sub-event goals were achieved. Results: During the study, 3361 telestrokes were completed and 306 (9.1%) patients received IV thrombolytics. After exclusions, 253 patients treated with IV thrombolytics were included. Five sub-events contributed to DTN time above and beyond the nuisance variables: door-to-telestroke request, stroke physician response time, telestroke phone-to-video, video duration prior to needle, and video completion-to-needle; each p <0.001. DTN time within 45 minutes was more likely when door-to-telestroke request <10 minutes (OR=12.30, 95%CI 3.47-43.65), video completion to needle <1 minute (OR=4.21, 95%CI 1.45-12.20) and telestroke phone-to-video <7 minutes (OR=5.24, 95%CI 1.41-19.49). Conclusions: Telestroke sub-events involving door-to-telestroke request, stroke physician response, telestroke phone-to-video, video duration prior to needle, and video completion-to-needle significantly contribute to DTN time. Successful achievement of sub-event goals was related to greater likelihood of administration of thrombolytic therapy within 45 minutes.