Vision-based localization plays a critical role in ensuring the positioning continuity of vehicles when Global Navigation Satellite System (GNSS) signals are unavailable. Although visual localization provides an effective auxiliary solution under GNSS-denied conditions, its performance is often constrained by insufficient landmarks. To address these limitations, collaborative vision-based localization via vehicle-to-vehicle (V2V) communication has been introduced, enabling vehicles to exchange positioning information and mitigate localization failures caused by landmark scarcity at individual nodes. However, existing studies predominantly emphasize algorithmic design, while a unified probabilistic framework for systematic performance analysis remains largely unexplored. To bridge this gap, this paper develops a novel analytical framework for collaborative vision-based localization in vehicular networks based on stochastic geometry. Specifically, the environmental landmark distribution is modeled using a homogeneous Poisson Point Process (HPPP), while vehicle locations are characterized by a Poisson Line Cox Process (PLCP). On this basis, we first derive the successful localization probability of a single vehicle relying solely on vision in GNSS-denied conditions. We then analyze the coverage probability of V2V transmissions under a Nakagami-$m$fading channel. Leveraging the derived coverage probability, a closed-form expression for the time-of-arrival (TOA)-based multi-vehicle collaborative localization probability is obtained. Finally, we define and characterize the overall GNSS-denied localization probability, which serves as a unified system-level metric quantifying the likelihood that an arbitrary vehicle can be successfully localized without GNSS support. The proposed framework explicitly reveals the coupled impacts of environmental uncertainty, wireless channel fading, and vehicular spatial distribution, thereby providing a theoretical benchmark for performance evaluation and parameter optimization of collaborative vision-based localization in vehicular networks.