Methodology for evaluating the robustness of multi-object tracking algorithms to input video stream degradation

Abstract

The paper presents an extended methodology for evaluating the robustness of multi-object tracking algorithms under degradation of video streams in intelligent video surveillance systems. The study is relevant because, in real conditions, video data are affected by spatial, photometric, temporal and observability-related distortions, such as blur, noise, compression artefacts, reduced resolution, frame loss, lower frame rate and occlusions. These degradations reduce detector stability, increase missed detections and false alarms, fragment trajectories and cause identity switches. The methodology uses a modular experimental pipeline comprising video-data preparation, controlled degradation modelling, object detection, tracker execution, ground-truth adaptation and quality-metric calculation. Fair comparison is ensured through factor isolation: in each run, one degradation parameter is changed, while other conditions remain fixed. The method combines multi-object tracking metrics with robustness indicators. Tracking effectiveness is assessed by Multiple Object Tracking Accuracy, Higher Order Tracking Accuracy, Identification F1 Score, Identity Switches, False Positives, False Negatives, trajectory fragmentation and processing speed in frames per second. For quantitative robustness analysis, the paper introduces the Robustness Score and Breaking Point indicators, which characterize absolute tracking quality and the rate of its decline as distortion intensity increases. The methodology supports comparative analysis of trackers, including Simple Online and Realtime Tracking, Deep Simple Online and Realtime Tracking, ByteTrack, Observation-Centric Simple Online and Realtime Tracking and CenterTrack, and helps select solutions for intelligent video surveillance systems under limited video-stream quality.