Generator-as-a-Matcher: Joint Optimization of Tracklet Matching and Gap Filling for Sparser mmWave Radar Placements on Smart Freeways

Millimeter-wave (mmWave) radar technology is increasingly used to collect vehicle trajectory data on smart freeways. However, the high cost of dense mmWave radar deployments has led to interest in sparser sensor placements to reduce infrastructure expenses or extend coverage. Sparse placements, however, introduce sensing gaps where vehicles may go undetected, resulting in identity loss within blind spots. This fragments trajectories into unlinked tracklets and complicates continuous vehicle detection and tracking. To address this challenge, we propose a deep generative learning approach that jointly matches vehicle identities over the blind-spot segments and reconstruct vehicle trajectories to fill the sensing gaps. Specifically, we introduce the Generator-as-a-Matcher framework, which co-optimizes a conditional Generative Adversarial Network (cGAN) with a bipartite graph matching method. By using latent motion features and a differentiable Sinkhorn approximation, the bipartite graph matching is integrated into the end-to-end training process with the trajectory generator, improving matching consistency during training. We evaluate our method using the highD, NGSIM, and TJRD mmWave radar-based trajectory datasets. Results indicate that tracklet matching and gap filling remain effective with mmWave radar sensor gaps of up to 200–300 m in free-flow traffic and up to 100 m in congested conditions. This corresponds to a potential reduction in sensor density by approximately 50–67% without substantial loss in tracking performance. The method operates solely on trajectory data and does not require video input, supporting its applicability for identity tracking using trajectory information alone.