Cross-Domain Multi-Semantic Trajectory Anomaly Detection via Kinematic Constraints and Bidirectional Sequential Reconstruction

Abstract

To address the bottlenecks of weak cross-domain generalization, insufficient representation of complex spatial behaviors, and the scarcity of real anomaly labels in the detection of multi-semantic pedestrian trajectory anomalies (e.g., running, loitering, and wrong-way) in public security, this paper proposes an unsupervised detection framework combining kinematic constraints and a Bidirectional Long Short-Term Memory Autoencoder (Bi-LSTM-AE). While deep graph neural networks and attention-based models capture multi-agent interactions, they are frequently prone to spatial weight drift and severe computational overhead when transferred across disparate surveillance layouts. In contrast, our framework focuses on the intrinsic kinematic manifold of individual trajectories, striking an effective trade-off between cross-domain generalizability and linear computational complexity O(L). To this end, we construct a rigorous cross-domain evaluation paradigm, utilizing normal trajectories from a massive public dataset (SDD) to drive the unsupervised training of the model, aiming to fit a generalized human walking manifold unconstrained by specific physical scenes. To tackle the lack of multi-dimensional evaluation benchmarks, we design a multi-semantic anomaly generation strategy based on underlying kinematic constraints, programmatically synthesizing highly realistic challenging test samples on the ground-truth manifold of the target domain (UCY). Cross-domain experimental results demonstrate that the proposed framework can overcome the Gaussian smoothing trap of generative models and the cross-domain generalization degradation of complex graph networks. In highly challenging unseen scenes, the overall detection F1 score is maintained above 0.90; notably, in the detection of “spatial loop anomalies (loitering)” that relies on a global spatio-temporal receptive field, a precise recall rate of 93.76% is achieved. With linear computational complexity, our method provides a novel solution for highly robust intelligent video surveillance early-warning in the open world.