Diffusion Model Learns City Traffic from Cameras Without Any Trajectory Labels

Vehicle trajectory data sit at the heart of modern intelligent transportation systems, feeding applications from signal optimization and flow analysis to accident investigation. Yet the raw material for training trajectory generators has always been awkward to obtain: GPS traces raise privacy and storage concerns, while surveillance cameras capture abundant but fragmented observations of vehicles passing discrete points. A study published in Computational Urban Science on 10 August 2026 proposes a way out of this bind. Researchers TongHao Liu and JiaoLing Zheng present a zero-shot trajectory generation method built on a diffusion model that never sees a single complete vehicle trajectory during training, relying instead on aggregated traffic flow statistics from surveillance cameras and public road network topology.

The term zero-shot here has a precise meaning: generating valid trajectories without any complete single-vehicle trajectory annotations. The authors draw a sharp boundary around their data usage. The diffusion model is initially pre-trained on randomly generated road network weight matrices, which are completely unlabeled. Real origin and destination data are used only for OD pair estimation, aggregated camera statistics serve as conditional inputs, and a small subset of real GPS trajectories is reserved exclusively for final performance validation, never entering training or optimization. This separation is what distinguishes the work from prior generative approaches that require continuous trajectory datasets for supervised training.

The technical core of the method is a reformulation of what gets generated. Rather than producing latitude and longitude sequences directly, the model generates weight distributions over road segments of the traffic network. Each vehicle receives a weight matrix that defines its navigation policy, and trajectories are then derived by applying the weighted A* path planning algorithm, which computes minimum-weight paths from each origin to each destination. Because every trajectory is a genuine path through the directed graph of the road network, topological validity is guaranteed by construction. The authors report a topological compliance score of 1.000 across all tested scenarios, meaning no generated trajectory contained wrong-way driving, prohibited turns, or restricted road access.

The diffusion model itself follows the standard denoising diffusion probabilistic framework, with a forward process that gradually adds Gaussian noise to clean weight matrices and a reverse process that learns to remove noise, conditioned on the complete time series of target macroscopic traffic states. Classifier-free guidance controls how strongly the conditioning constrains generation: values above 1.0 push trajectories closer to the real distribution at some cost to diversity, while lower values do the opposite. A hybrid autoencoder with spatial convolution compresses the high-dimensional road weight matrices into a compact latent space while preserving topological connectivity, and residual MLP blocks fuse the encoded flow features with the latent weight features.

Generation alone is not enough, so the team wraps the diffusion model in an adaptive black-box optimization loop. Candidate weight matrices are sampled from the diffusion model, evaluated by simulating trajectories and computing the divergence, measured as Jensen-Shannon divergence summed over time, between simulated and target macroscopic states. The top-scoring candidates augment the training dataset, and the diffusion model is retrained to focus on promising regions of the weight space. The authors frame the discrepancy between simulated macroscopic traffic states and target aggregated flow statistics as a self-supervised feedback signal, making the whole pipeline a two-stage self-supervised learning mechanism that needs no real trajectory supervision at any point.

Experiments were conducted on real-world traffic data from Dadukou District in Chongqing, an urban road network with 114 directed edges and 573 intersections. Surveillance camera data were collected from 7:00 to 9:00 AM on August 21, 2023, capturing 1,557 vehicles at discrete observation points. A practical hurdle was estimating individual origin-destination pairs from partial camera observations. Vehicles seen at multiple cameras get their origin and destination from the first and last observations, while single-camera vehicles, which account for 38.7 percent of detections in Dadukou, have destinations sampled from a historical exit probability matrix built from three months of aggregated traffic data. Validated against 127 GPS ground-truth vehicles, OD estimation reached 89.2 percent accuracy for multi-camera vehicles and 76.5 percent for single-camera vehicles.

The results compare favorably against four baselines adapted for the sparse setting: DGD, Traj-VAE, RL-diff, and MoveSim. In the demanding 1000-vehicle scenario, the proposed method achieved a travel time distribution similarity of 0.061, a 66.5 percent reduction over the best baseline, RL-diff at 0.182. Turning behavior distribution similarity stayed below 0.05 across all scenarios. On the macroscopic side, the method recorded a Jensen-Shannon divergence of 0.049 for 100 vehicles, a 23.4 percent improvement over Traj-VAE’s 0.064, and 0.480 versus Traj-VAE’s 0.650 at 1000 vehicles. It also achieved the highest unique path ratio in every scenario, with the smallest decline as traffic volume scaled from 100 to 1000 vehicles, indicating strong diversity retention in large-scale generation.

An ablation study confirmed that both key components matter. Removing the autoencoder and optimizing directly in high-dimensional weight space dropped the topological compliance score from 1.000 to 0.856 under 1000-vehicle conditions and raised Jensen-Shannon divergence substantially at every traffic density. Disabling classifier-free guidance increased divergence by 22.4 percent for 100 vehicles and 43.5 percent for 1000 vehicles, and reduced the unique path ratio from 78.2 percent to 71.3 percent. The method also transferred to an independent dataset from Jiulongpo District with a different road network: the pre-trained diffusion model was moved over without re-pre-training, the full optimization cycle was rerun using only Jiulongpo’s flow statistics and topology, and hyperparameters were left untouched, with the method maintaining perfect topological compliance while baselines showed severe inconsistencies on the unseen network.

Efficiency figures suggest practical deployment is feasible. The optimization converges after roughly 50 iterations, and for the 1000-vehicle scenario the average per-iteration runtime is 15 seconds on a single NVIDIA RTX 3090 GPU, giving a total training time of about 10 minutes with peak GPU memory usage of 2 GB. Runtime scales linearly with the number of vehicles and road segments, and the method achieved a 4.6-fold speedup over RL-diff and a 2.1-fold speedup over DGD at 1000 vehicles. The authors note a limitation on transparency: the datasets derive from a commercial project under confidentiality agreements and cannot be released, though source code and experimental protocols are available from the corresponding author on reasonable request. As long as aggregated flow statistics exist, the authors state, the method applies to any road network representable as a directed graph, positioning it as a template for generating simulation-ready, road-constrained trajectories in cities where complete trajectory data simply do not exist.

Subject of Research: A zero-shot vehicle trajectory generation method that uses a self-supervised diffusion model trained only on aggregated surveillance traffic statistics and road network topology.

Article Title: Zero-shot trajectory generation with self-supervised diffusion model

Article References: Liu, T., & Zheng, J. (2026). Zero-shot trajectory generation with self-supervised diffusion model. Computational Urban Science, 6(1), Article 51. https://doi.org/10.1007/s43762-026-00280-7

Image Credits: AI Generated

DOI: 10.1007/s43762-026-00280-7

Keywords: trajectory generation, diffusion model, self-supervised learning, intelligent transportation systems, surveillance cameras, road networks, zero-shot learning, traffic simulation, generative models, urban traffic, Zero-shot, trajectory