Scale-invariant mask-guided vehicle keypoint detection from a monocular image

Abstract

Intelligent vehicle detection and localization are important for autonomous driving systems, particularly traffic scene understanding. Robust vision-based vehicle localization directly affects the accuracy of self-driving systems but remains challenging to implement reliably due to differences in vehicle sizes, illumination changes, background clutter, and partial occlusion. Bottom-up-based vehicle detection using vehicle keypoint localization efficiently provides semantic information for partial occlusion and complex poses. However, bottom-up-based approaches still struggle to handle robust heatmap estimation from vehicles with scale variations and background ambiguities. This paper addresses the problem of predicting multiple vehicle locations by learning semantic vehicle keypoints using a multi-resolution feature extractor, an offset regression branch, and a heatmap regression branch network. The proposed pipeline estimates the vehicle keypoint by effectively eliminating similar background features using a mask-guided heatmap regression branch and emphasizing scale-adaptive heatmap features in the network. Quantitative and qualitative analyses, including ablation tests, verify that the proposed method is universally applicable, unlike previous approaches.