แนวคิดหลัก
The critical role of pedestrian bounding box, ego-vehicle speed, and local context features in predicting pedestrian crossing intentions, with body pose being less significant. The analysis reveals potential biases introduced by the speed feature and proposes an alternative feature representation to mitigate this.
บทคัดย่อ
The study evaluates the performance of five distinct deep neural network models for predicting pedestrian crossing intentions using the Pedestrian Intention Estimation (PIE) dataset. The models are assessed across various contextual scenarios, including roadway type, traffic light state, crosswalk state, proximity to the ego-vehicle, and ego-vehicle speed.
The key findings are:
-
Context-aware Performance Evaluation:
- The models exhibit nuanced differences in performance across the various contextual scenarios.
- Midblock crossing scenarios pose the greatest challenge, resulting in the lowest performance.
-
Feature Importance Analysis:
- The pedestrian bounding box is the most important feature, followed by ego-vehicle speed and local context features.
- Body pose is deemed less significant, potentially due to susceptibility to noise and occlusion.
- The models exhibit a striking resemblance in how they respond to feature permutations, suggesting the fundamental relevance of certain features.
-
Ego-vehicle Motion Feature:
- The speed feature can introduce bias by capturing ego-vehicle behavior rather than pedestrian behavior.
- An alternative feature representation, the pedestrian-vehicle proximity change rate, is proposed to mitigate this bias, but does not yield significant performance improvements.
The study highlights the importance of considering contextual factors and diverse feature representations in developing accurate and robust intent-predictive models for pedestrian crossing scenarios. Future research should focus on addressing challenges in complex traffic environments and exploring novel feature representations to enhance predictive capabilities and pedestrian safety.
สถิติ
The pedestrian bounding box feature contributes 9.1% (σ=1.22) to accuracy, 9.2% (σ=1.2) to AUC, and 9.1% (σ=1.23) to the F1 score, achieving the highest importance scores across all models and scenarios.
The ego-vehicle speed feature contributes 5.1% (σ=2.11) to accuracy, 5% (σ=2.06) to AUC, and 5.1% (σ=2.1) to F1 score.
The local context feature contributes 4.7% (σ=0.71) to accuracy, 4.6% (σ=0.73) to AUC, and 4.7% (σ=0.76) to F1 score.
The body pose feature contributes 1.3% (σ=0.46) to accuracy, 1.4% (σ=0.47) to AUC, and 1.3% (σ=0.46) to F1 score.
คำพูด
"The critical role of pedestrian bounding box, ego-vehicle speed, and local context features in predicting pedestrian crossing intentions, with body pose being less significant."
"The speed feature can introduce bias by capturing ego-vehicle behavior rather than pedestrian behavior."