A longitudinal scanline based vehicle trajectory reconstruction method for high-angle traffic video

Abstract

In this paper, a robust and efficient High-angle Spatial-Temporal Diagram Analysis (HASDA) model is built to reconstruct high-resolution vehicle trajectories from infrastructure traffic surveillance videos. A combined methodology was developed, comprising of scanline-based trajectory extraction and feature-matching coordinate transformation. A scanline-based trajectory extraction technique is introduced to separate vehicle strands from pavement background on the spatial-temporal diagram by considering color features, gradient features, and motion features. Particular cleaning algorithms for removing static object noises, shadows, and occlusions are also established. Feature-matching coordinate transformation converts the pixel coordinates to the real-world coordinates to generate the physical vehicle trajectory. To evaluate the algorithm, generated trajectory results were compared to the reconstructed version of the Next Generation Simulation (NGSIM) dataset. 15-min NGSIM video was divided into a 5-min dataset for the calibration and the remaining 10-min data for evaluation. Model parameters calibrated based on the 5-min video data are then applied to the 10-min testing data. Two levels of performance measurements are considered to evaluate both trajectory-level and point-level results. A reference algorithm based on mainstream motion-based detection and tracking methods are used as a baseline algorithm. Based on the evaluation results, the proposed method shows promising trajectory detection results, that on average more than 90% of vehicle trajectories are constructed by the proposed methods from the NGSIM videos. The HASDA model outperforms the reference algorithm and shows superior transferability in the training-testing experiment. Further work needs to be done to improve the algorithm performance against shadows and occlusions by incorporating more intelligent and advanced techniques.