Anti-occlusion Tracking Algorithm for Bird Flock Based on Spatiotemporal Topology

To address the challenges of appearance similarity, weak visual features, and frequent occlusion in bird flock imagery, this paper proposes an anti-occlusion tracking algorithm based on spatiotemporal topological feature. A binary spatiotemporal descriptor is constructed to encode the relative positional feature between individuals and their neighbours, overcoming the limitations of appearance-based features. An occlusion reasoning module integrated with multi-view geometric constraint is designed. It adopts a sliding window trajectory enumeration and optimal assignment strategy, effectively resolving the issue of trajectory identity ambiguity caused by occlusion. The results indicate that the proposed descriptor achieves a discriminability (Kullback-Leibler divergence) of 2.18 ~ 3.95, outperforming traditional and deep features. The proposed 2D tracking algorithm attains a multi-object tracking accuracy (MOTA) of 93.5% ~ 96.2%, while the 3D tracking algorithm elevates MOTA to 98.3%.