Enhancing Vehicle Re-identification and Matching for Weaving Analysis

TL;DR

This work tackles the challenge of quantifying lane-level vehicle weaving on highways by leveraging two non-overlapping cameras positioned at the weaving transition points P1 and P2. It introduces a two-phase framework: (1) constructing a dedicated weaving ReID dataset (4,902 unique vehicles, 78,979 images) from nine weaving areas using highway and drone cameras, and (2) training a Vision Transformer–based ReID model (vit_base_patch16_224) with L_id and L_triplet supervision, followed by cross-area matching via a modified Hungarian algorithm that blends feature similarity and time alignment relative to the average travel time $T_a$. The approach achieves a mean average precision of 47.8% and strong rank-based retrieval (Rank-1: 42%, Rank-5: 50.9%, Rank-10: 57.2%), with Grad-CAM visualizations confirming the model attends to discriminative vehicle regions. By treating the matched subset as a representative sample, the method enables estimation of lane-specific weaving patterns despite non-overlapping camera views, offering a practical data-driven tool for traffic management and infrastructure planning.

Abstract

Vehicle weaving on highways contributes to traffic congestion, raises safety issues, and underscores the need for sophisticated traffic management systems. Current tools are inadequate in offering precise and comprehensive data on lane-specific weaving patterns. This paper introduces an innovative method for collecting non-overlapping video data in weaving zones, enabling the generation of quantitative insights into lane-specific weaving behaviors. Our experimental results confirm the efficacy of this approach, delivering critical data that can assist transportation authorities in enhancing traffic control and roadway infrastructure.

Figures (7)

• Figure 1: Weaving examples. Vehicles at P1 can either come from the ramp or the highway. As vehicles move to P2, they can stay on the highway or exit to the ramp.
• Figure 2: The framework of weaving analysis consists of two phases:Weaving Dataset and Weaving ReID Model and Weaving Analysis. (Left) Phase 1: We create a custom weaving ReID dataset by matching vehicles across nine specific weaving areas. Subsequently, we train our custom weaving ReID model on this dataset. This phase focuses on establishing the groundwork for vehicle identification and tracking. (Right) Phase 2: We match vehicles across weaving areas using the ReID model, extracting distinctive features from vehicle images. These features, combined with spatial-temporal information, are used in our Hungarian Matching module to derive lane-specific weaving patterns for comprehensive analysis. More details are explained in Section \ref{['sec:method1']} and \ref{['sec:method2']}.
• Figure 3: Some examples extracted from our Weaving ReID dataset. All the images keep their original size, shape and aspect ratio. For each unique vehicle, it has about 15 image samples.
• Figure 4: Videos from five weaving areas are used in our experiments.
• Figure 5: Feature Similarity Analysis involves extracting feature vectors for query, positive, and negative vehicle samples from our ReID model. The cosine similarity between positive pairs is significantly higher than that between negative pairs, providing a robust criterion for vehicle matching.