Open Autonomous Driving Scene Object Detection Based on MEBA-OWOD Algorithm
WANG Chao
SU Shuzhi
ZHU Yanmin
XU Yang
Abstract：To address the complex object occlusion,dense object,and small object miss-detection issues in autonomous driving scenarios,an open-world object detection algorithm based on multi-scale feature enhancement and boundary awareness(MEBA-OWOD)was proposed.The algorithm first constructed a multi-scale feature enhancement(MSFE)module to meticulously filter and emphasize the extracted multi-scale image features,aiding the model in identifying and focusing on important and critical details in the images.Then,a boundary aware redundant box filter(BARF)was designed,which effectively filtered overlapping boxes of known objects and redundant boxes of unknown objects,reducing the occurrence of proposal boxes containing incomplete objects and dense objects.The experiments conducted on the way scenes(Wayce)autonomous driving dataset demonstrated that,compared to several excellent object detection algorithms and open world object detection algorithms,the MEBA-OWOD algorithm was increased by 0.3％and 49.6％respectively on the indicators of known class mean average precision and unknown class average precision companed with the second-best comparison algorithms.This algorithm can enhance the detection of unknown objects without affecting the detection of known objects,and can be effectively applied in the field of autonomous driving.
Keywords：open world object detectionautonomous drivingboundary awaremulti-scale feature enhancementfeature selectionredundancy filtering
Publication Date：2024-12-20
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：7( 514-520 )