ROI-WEIGHTED PRIMARY FRAME FILTERING FOR EFFICIENT DEEP VIDEO SURVEILLANCE

Abstract

Real-time video surveillance systems frequently rely on deep neural detectors and multi-object trackers.
Although these models provide high detection and tracking quality, processing every frame is computationally expensive,
especially for long-term multi-camera deployments. This paper presents an ROI-weighted primary filtering algorithm that
decides whether an incoming video frame should be processed by a full deep model or skipped and propagated using
previously computed results. The method is motivated by the observation that ordinary mean absolute frame difference
treats all pixels equally. At the same time, surveillance decisions are usually more sensitive to changes in object regions
than to background fluctuations. The proposed estimator constructs a spatial importance map from previously detected
bounding boxes, expands these regions by a dilation margin, and computes a normalized weighted frame-difference
score. A cost-minimization rule converts this score into a binary processing gate. Experiments on MOT20 sequences
show that the ROI-weighted score amplifies object-related changes by 1.15x to 1.87x relative to global MAFD. Practical
deployment scenarios demonstrate that more than 65% of frames can be handled by lightweight filtering, reducing computational
load by 56.69% and 61.31% in two surveillance cases. These findings indicate that object-aware frame filtering
is a simple and effective pre-inference mechanism for resource-efficient video analytics.