Real-time Pedestrian Detection Using LIDARandConvolutional Neural Networks

TL;DR: The evaluation results indicate that the use of the LIDAR-based ROI detector can reduce the number of false positives by a factor of 2 and reduce the processing time by a Factor 4, which is above 90% when there is 1 false positive per second.

Abstract: Thispaperpresents anovel real-time pedestrian detection system utilizing aLIDAR-based object detector andconvolutional neural network(CNN)-based imageclassifier. Ourmethodachieves over10 frames/second processing speedbyconstraining thesearch spaceusing therangeinformation fromtheLIDAR.Theimageregion candidates detected bytheLIDARareconfirmed forthepresence ofpedestrians bya convolutional neural network classifier. OurCNN classifier achieves high accuracy atalowcomputational costthanks toitsability toautomatically learn asmall numberofhighly discriminating features. Thefocus ofthis paperistheevaluation oftheeffect ofregion ofinterest (ROI)detection onsystem accuracy andprocessing speed. Theevaluation results indicate thattheuseoftheLIDAR-based ROIdetector canreduce thenumber offalse positives byafactor of2andreduce theprocessing timebya factor of4.Thesingle framedetection accuracy ofthesystem isabove 90%whenthere is1false positive persecond. I.INTRODUCTION Morethan3000pedestrians arekilled eachyearintraffic accidents inJapan. Looking atthereason ofthese accidents, itisalmost always thelackofattention ontheside ofthedriver. Similar statistics have beenreported inother countries aswell. Therehasbeenagreat dealofinterest inrecent years inthedevelopment ofpedestrian detection systems that could helpreduce thenumberandimpact of these accidents. Mostoftheproposed systems useacameraasthe sensor, because cameras canprovide thehighresolution needed for accurate classification andposition measurement. Cameras canalso beshared withother safety support subsystems inthecar, suchas alane-keep assist system, improving theprice-benefit ratio ofthe camera. Thedisadvantage ofimage-only detection systems isthehigh computational costassociated withclassifying alarge numberof candidate imageregions. Accordingly, ithasbeenatrend forseveral years touseahierarchical detection structure combining different sensors. Inthefirst steplowcomputational costsensors identify asmallnumberofcandidate regions ofinterest (ROI). Sincethe accuracy ofthese faster sensors islimited, thedetected ROI-sneed tobeconfirmed withmoreaccurate sensors. Theadvantage ofsuch ahierarchical structure isthat thehighaccuracy sensors needtobe applied only toasmall numberofcandidates, therefore therequired computation ismuchless than inamonolithic system. Thetradeoff is that pedestrians notdetected bythefaster sensor cannot bedetected bylater sensors either, limiting theoverall detection rateofthe system. A.PaperOutline