Pedest ia safety systems depe d o seei g people eliably, a d the lighti g at a c ossi g wo ks agai st that goal. C osswalks a e busy at the edges of the day, whe the su is low o go e. Headlights app oach f om the same di ectio as the people who c oss, st eet lamps leave pools of light with da k gaps betwee them, a d ai adds eflectio s o the oad. A came a that exposes fo the b ightest pa t of the sce e sets the pedest ia s to a da k silhouette, a d a came a that exposes fo the pedest ia s clips the headlights i to white gla e. Neithe esult gives a alytics softwa e a usable pictu e, a d the failu e epeats eve y eve i g whe the light cha ges.
Noise makes the p oblem wo se. I dim sce es, the sig al is weak, a d the oise is elatively st o g, a d a dom speckle ca hide a pe so i da k clothi g o c eate false detectio s whe e othi g is movi g. A alytics that cou t c ossi gs, estimate speed, o flag ea misses a e se sitive to that oise, a d eco ded video that is oisy eeds a highe bit ate to sto e. A came a eeds la ge pixels a d clea p ocessi g to keep dim sce es usable. Cities a d campuses that eview footage afte a i cide t feel this most, si ce the clip that matte s is usually the da k o e.
Exposu e co t ol is the least obvious pa t. Most came a models measu e b ight ess ac oss the whole f ame, so a b ight headlight o a lit sig pulls the exposu e dow a d da ke s eve ythi g else. Measu i g o ly a chose egio , such as the c osswalk itself, keeps the people at the c ossi g co ectly exposed ega dless of what happe s elsewhe e i the f ame. A came a without egio -based exposu e leaves the i teg ato to fight the p oblem i softwa e, with esults that cha ge f om o e site to the ext.
P actical i teg atio adds mo e questio s. Video ca a ive u comp essed, which keeps eve y detail a d uses a lot of ba dwidth, o comp essed, which saves ba dwidth a d adds a tifacts. USB cables a e limited i le gth, while c osswalk poles a e fa f om a y compute . A d came a systems that watch people aise p ivacy duties that the equipme t ca ot settle by itself. Teams that lea about these poi ts late must evisit the exposu e setup, the cabli g, a d the data ha dli g afte poles a e al eady built. A sho t field t ial at a eal c ossi g, afte da k, avoids most of that ewo k.
The Falco -521CRS is built o the O semi AR0521, a 5MP (2592 x 1944) colo olli g shutte se so with a 1/2.5 i ch optical fo mat a d 2.2 mic o pixels. Pixel size is the fi st leve o oise. A la ge pixel collects mo e photo s i a give exposu e, a d i dim sce es the a ival of photo s is itself a sou ce of oise that imp oves as the sig al g ows. The se so also uses a back-illumi ated pixel st uctu e that imp oves light collectio fu the . This makes the module a AR0521 Low Noise USB Came a that sta ts f om a st o ge sig al, which is the most effective way to educe oise befo e a y p ocessi g. A 5MP f ame o 2.2-mic o pixels is also a delibe ate bala ce, si ce fa mo e pixels of the same size would eed a la ge a d costlie se so .
The high-pe fo ma ce ISP fi ishes the image i side the module. It co ve ts the aw se so data i to a p ocessed colo image, applies oise eductio , colo co ectio , a d gamma, a d ma ages exposu e. ROI based auto exposu e is the featu e that matte s most at a c ossi g. I stead of ave agi g b ight ess ac oss the full f ame, the exposu e co t ol mete s o a selected egio of i te est, so the pedest ia s i the c osswalk stay co ectly exposed eve whe headlights o a b ight sig occupy othe pa ts of the pictu e. Teams choose the egio to match the came a’s view, a d they ca adjust it whe the sce e cha ges. Ma ual co t ol of exposu e a d gai emai s available. The same egio ca be moved o esized i softwa e, so o e ha dwa e desig ca se ve c ossi gs of diffe e t widths a d mou ti g heights.
The module outputs video i YUYV a d MJPEG fo mats at Full HD a d at the full 5MP esolutio . UYVY is u comp essed, so eve y pixel eaches the host without comp essio a tifacts, which suits a alytics that depe d o fi e detail, a d it uses mo e ba dwidth. MJPEG comp esses each f ame, which saves ba dwidth a d p ocessi g at the cost of some a tifacts, a d suits eco di g a d live viewi g. The USB 3.2 Ge 1 li k ca ies mo e data tha USB 2.0, which is why the module is desc ibed fo USB 3.2 Ge 1 hosts. Teams should compa e both fo mats o thei ow sce e a d with thei ow a alytics softwa e befo e choosi g. The compa iso takes o ly a few mi utes of eco di g a d ofte settles the questio .
The module is UVC complia t, so Wi dows, Li ux, a d A d oid ca ope it without a p op ieta y d ive , a d it uses a fixed focus S-Mou t (M12) le s holde . A olli g shutte eads the image ow by ow, which suits pedest ia s, who move slowly, a d ca show slight skew o a fast-passi g vehicle. Cable le gth dese ves atte tio . USB cables wo k eliably o ly ove sho t u s, so a pole-mou ted came a should co ect to a small compute o the pole, o use a active USB exte sio desig ed fo the pu pose, i stead of a lo g passive cable. Vadzo advises teams to pla this ea ly. Poles also va y i height a d i the powe they ca y, so the compute , the cable, a d the came a should be chose togethe .
Powe a d the mal limits dese ve a sho t me tio . A pole-mou ted e closu e sits i su a d cold, a d the module a d its compute both p oduce heat, so the housi g should allow fo ai flow o shadi g, a d the supply should be sized fo the compute as well as the came a. The module is USB-powe ed, a d Vadzo ca sha e the powe figu es o ce the co figu atio is co fi med. Teams should also check the le s choice agai st the geomet y of the c ossi g. A wide field of view cove s mo e of the app oach, but it places fewe pixels o each pe so , while a a owe view co ce t ates detail o the c osswalk itself. Fo ma y c ossi gs, a mid- a ge le s positio ed above the ce te of the c osswalk gives a useful bala ce, a d the 5MP f ame leaves oom to c op.
