Beyond Overall Accuracy: Pose- and Occlusion-driven Fairness Analysis in Pedestrian Detection for Autonomous Driving

TL;DR

The paper introduces a pose- and occlusion-aware fairness framework for pedestrian detection in autonomous driving, labeling four pose attributes and instance size, and evaluating eight detectors on the ECP-DP dataset. It quantifies biases with Equal Opportunity Difference and Cohen’s $h$ across multiple confidence thresholds, backed by a $Z$-test for significance. Key findings reveal stronger miss-rate biases due to lower-body occlusions and lateral views, with Cascade R-CNN delivering the best overall fairness and several detectors improving fairness after targeted fine-tuning. The work demonstrates that achieving equitable safety requires explicit consideration of pose, occlusion, and view-dependent biases in detector design and evaluation, offering a scalable methodology for fairness analysis in AD perception.

Abstract

Pedestrian detection plays a critical role in autonomous driving (AD), where ensuring safety and reliability is important. While many detection models aim to reduce miss-rates and handle challenges such as occlusion and long-range recognition, fairness remains an underexplored yet equally important concern. In this work, we systematically investigate how variations in the pedestrian pose -- including leg status, elbow status, and body orientation -- as well as individual joint occlusions, affect detection performance. We evaluate five pedestrian-specific detectors (F2DNet, MGAN, ALFNet, CSP, and Cascade R-CNN) alongside three general-purpose models (YOLOv12 variants) on the EuroCity Persons Dense Pose (ECP-DP) dataset. Fairness is quantified using the Equal Opportunity Difference (EOD) metric across various confidence thresholds. To assess statistical significance and robustness, we apply the Z-test. Our findings highlight biases against pedestrians with parallel legs, straight elbows, and lateral views. Occlusion of lower body joints has a more negative impact on the detection rate compared to the upper body and head. Cascade R-CNN achieves the lowest overall miss-rate and exhibits the smallest bias across all attributes. To the best of our knowledge, this is the first comprehensive pose- and occlusion-aware fairness evaluation in pedestrian detection for AD.

Figures (5)

• Figure 1: Sample scene image from the ECP-DP braun21simple, annotated with joints. Yellow joints are visible, and orange ones are occluded. Green boxes show correct detections, and red ones show missed pedestrians (by YOLOv12-S yolov12). Automatically extracted pose-related attributes appear in white text.
• Figure 2: Pose- and Occlusion-Driven Fairness Analysis Framework. Top: ECP-DP braun21simple images are processed by eight detectors and matched to ground truths via IoU to compute true positives and false negatives per subgroup. Bottom: Pose attributes (leg status, elbow status, body orientation) and occlusion flags for 17 joints are extracted from annotations. Subgroup detection results are then evaluated with the fairness metric to quantify bias.
• Figure 3: Illustration of random instances from pedestrian legs status, elbow status, and body orientation.
• Figure 4: Per-joint occlusion rates (percentage) across body keypoints in the merged ECP-DP training/validation set. The red dashed line indicates the average occlusion rate (39.3%) across all joints.
• Figure 5: Visual comparison of all detectors across attributes and joint occlusion ($\tau=0.25$). For body orientation, three EOD values are averaged. EOD values for left and right joints are also averaged due to similar values.