<p>Utility tunnels represent a critical component of urban infrastructure, often regarded as the city’s lifeline. Ensuring the security of utility tunnels is crucial for protecting urban infrastructure and preventing threats to essential services from intruders. Traditional trajectory recovery methods rely on GPS data, which are not suitable for underground environments with inferior signals. The widely deployed network of surveillance cameras in tunnels presents new opportunities, as it allows us to extract video signals and recover intruders’ trajectories, but this approach encounters several challenges. Firstly, the videos from camera network lacks clear identity labels, leading to ambiguous correspondence between intruders and video records. Secondly, the uneven and incomplete distribution of surveillance cameras results in data sparsity issues. In this paper, we propose a novel two-stage framework, TrajUT, which leverages machine learning algorithms to address these challenges. In the first stage, we introduce the multi-criteria-based trajectory segmentation to segment the raw record sequence into sub-sequences belonging to different groups of intruders, based on the continuity of sequences. For second stage, we design a hidden Markov model-based method that uses the records as observed sequences to infer the hidden sequences with the minimum cost, which correspond to the intruder movement trajectories. The proposed dynamic speed constraints enhance trajectory recovery by incorporating both spatial distances and temporal speed relationships into the cost adjustment mechanism. We conduct experiments using real-world data from the utility tunnels in Suzhou City. The results demonstrate the effectiveness of our framework, achieving a 10.38% improvement in precision compared to the best baseline method.</p>