Anchor-Clustering: A Lightweight Efficient Method to Aggregate Human Attribute Prediction in Video Surveillance

We propose a two-stage, reliability-aware pipeline for video-based human attribute recognition. A pretrained vision–language model extracts attribute-relevant frame features, which a lightweight multilayer perceptron decodes into per-frame probabilities. To aggregate noisy frame scores into stable track-level outputs, we introduce the anchor-clustering aggregation method: probabilities are discretized into ordered anchors, and the final decision is based on anchor counts and the longest high-anchor streak, favoring temporal consistency over isolated confidence spikes. The proposed method yields relative improvements of 9.55 and 13.54% in True Positive Rate on the MARS-Attribute and DukeMTMC-VideoReID-Attribute datasets, respectively, while maintaining a high F1 score and minimal inference overhead of milliseconds. The proposed framework is also flexible and easy to adapt beyond human attribute recognition, making it applicable to a wide range of temporal aggregation tasks.