通道特征蒸馏的复杂道路场景目标检测

打开文本图片集

关键词：复杂道路场景；小目标；特征蒸馏；目标检测；两阶段蒸馏策略DOI：10.15938/j. jhust.2025.06.002中图分类号：TP391.41 文献标志码：A 文章编号：1007-2683（2025）06-0009-10

Abstract：Aiming atthe problem of denselydistributed smalltargetsandserious target occlusion incomplex roadscenes，asmall targetdetectionmethdbasedonchaelfeaturedistilltionisproposedInodertosoleteproblemofsalltargetseingliitedby resolutionandmulti-scalefeatureimbalance，asmallconvolutionkerelisusedintheNeckpart，andachannelatentionfeature enhancementmoduleisdesignedtoeliminatethedimensionalinfluenceinmulti-scalefeaturemapsandgeneraterichcontiuous features.Featuremapingenhancestheexpresionabiltyoffeaturesandsignificantlyimprovesthedetectionperformanceofsmal targets.Consideringteproblemsoffeaturelossndpattesplitinginoccludedtargets，twosagedistllaionstrategyisdopted First，strong Mosaicdataenancementisusedtogudethestudentmodeltobeterextractoluded argetinformation；ten，without Mosaicenhancement，thestudentmodelself-adjuststoimproveherobustness tooludedtargets.Duringthedisilatioprocess， softmax normalizationisperformedonthefeaturechanneldimensiontoconvertheactivationvalueintoaprobabilitydistribution.By minimizing theasymmetricKLdivergencebetweentheteacherandstudentchanelprobabilitydistributions，thestudentmodelpays moreattentiontoteforegoudaandacheveseteresults，chevingegraedfatureiteractionExpermetalrsultso that the distilled RTMDet-m reaches 73. 2% mAP and 93.8% mAP50 on the KITTI data set， with only 24.7 M parameters and 39.3 G FLOPs，significantlysurpassing the YOLOseriesandYOLOXseries，and theaverage acuracyofsmalltargetsiscomparable. Compared with the original model，it is improved by more than 3% ，meeting the vehicle and pedestrian detection requirements in complex road scenes.

Keywords：complex road scenes；small targets； feature distilation；object detection;two-stage distilltion strategy

0 引言

道路车辆和行人检测一直是自动驾驶、交通管理和智能交通系统等领域中的一个至关重要的任务[1-3]。（剩余16740字）