PEDESTRIAN INTENTION PREDICTION ON JAAD AND PSI-2: A COMPARATIVE STUDY OF RECURRENT AND ATTENTION-BASED MODELS UNDER CLASS IMBALANCE AND THRESHOLD SELECTION

Engineering notes

For consistency across datasets, all experiments use a unified 30-frame grouped-split sequence-classification protocol rather than the official JAAD or PSI-2 benchmark protocols.The study evaluates LSTM, BiLSTM, Attention-BiLSTM, and Dual-Attention BiLSTM models on fixed-length temporal sequence representations derived from JAAD and PSI-2.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Pedestrian intention prediction is important in intelligent transportation systems and autonomous driving because safe vehicle behavior depends not only on detecting pedestrians, but also on anticipating their likely future actions. However, performance in this task is strongly influenced by dataset properties, class imbalance, threshold selection, and evaluation design. This thesis investigates pedestrian intention prediction on the JAAD and PSI-2 datasets using recurrent and attention-based sequence models within a comparative methodology-focused framework. For consistency across datasets, all experiments use a unified 30-frame grouped-split sequence-classification protocol rather than the official JAAD or PSI-2 benchmark protocols.The study evaluates LSTM, BiLSTM, Attention-BiLSTM, and Dual-Attention BiLSTM models on fixed-length temporal sequence representations derived from JAAD and PSI-2. To support reliable interpretation, the thesis employs grouped train–test splits, repeated-seed evaluation, class-sensitive metrics, a controlled PSI-2 comparison of oversampling and class-weighted binary cross-entropy (BCE), threshold-policy analysis, and feature-group ablation. The results show that, under the unified experimental framework, JAAD exhibits comparatively stable model behavior, whereas PSI-2 is much more sensitive to model design, rebalancing strategy, and operating-point selection. The strongest observed mean balanced accuracy on PSI-2 was achieved by Attention-BiLSTM in the no-oversampling class-weighted setting, while class weighting provided no additional benefit under oversampling. Threshold-policy analysis showed that different operating objectives can change the preferred PSI-2 model, and feature-group ablation identified temporal information as the most important component of the PSI 19-dimensional representation. The main contribution of the thesis is methodological: in small and imbalance-sensitive datasets such as PSI-2, reliable conclusions depend not only on model architecture, but also on repeated-seed evaluation, rebalancing strategy, threshold-policy selection, and feature-level analysis.

Links and sources

Comments