Jetso AGX O i Came a is built fo dema di g pe ceptio wo kloads. It combi es a GPU, deep lea i g accele ato s, a d seve al came a i puts, so a si gle boa d ca u avigatio , detectio , a d eco di g fo a obot o a etail system at o ce. The quality of what it pe ceives still depe ds o the came a, a d two p ope ties of the came a cause most of the t ouble i p actice. The fi st is how the se so exposes the image. The seco d is whe e the image gets p ocessed. Both a e decided whe the came a is chose , lo g befo e the softwa e team sees a f ame.
Most se so s use a olli g shutte , which eads the image ow by ow. Whe the subject o the came a moves, the ows captu e diffe e t mome ts, a d st aight li es be d. A fo klift ca yi g a pallet, a obot a m sweepi g past a shelf, a shoppe walki g th ough a doo way, a d a came a o a movi g vehicle all c eate this effect. Fo a alytics that measu e positio s, ead codes, o t ack people, the disto tio matte s, because geomet y de ived f om a skewed image is w o g. Fast exposu es educe the skew but eed st o g light, which wa ehouses a d sto es do ot always p ovide. A global shutte avoids the issue by exposi g eve y pixel at the same i sta t, so movi g subjects keep thei shape. The be efit is g eatest whe subjects move quickly o whe the came a itself is mou ted o a movi g platfo m.
Image p ocessi g aises the seco d p oblem. A aw se so se ds u p ocessed Baye data to the Jetso , which must demosaic, co ect colo , educe oise, a d ma age exposu e. O Jetso , that wo k is do e by the host ISP, which eeds a tu i g p ofile fo each se so a d le s, a d good tu i g takes time a d expe tise. Two p oducts usi g the same Jetso ca look quite diffe e t because of tu i g alo e. A module that fi ishes the image itself emoves that task, but a buye must the u de sta d that the host ISP is ot pa t of the pipeli e. That is a fai excha ge fo ma y teams, a d a u welcome su p ise fo the few who expected aw access.
Co ecto s add a p actical questio . MIPI came a modules co ect th ough a defi ed umbe of data la es, a d ca ie boa ds diffe i the co ecto s they offe . A came a that suppo ts o ly o e la e cou t limits whe e it ca be used. Teams that discove these co st ai ts late must edesig the ca ie o cha ge the came a afte ha dwa e is built. A sho t eview of la e cou ts, co ecto s, a d image p ocessi g eeds at the sta t of a p oject avoids most of that ewo k.
The Bolt-234CGS is built o the O semi AR0234, a 2MP (1920 x 1200) colo global shutte se so with a 1/2.6 i ch optical fo mat a d 3.0 mic o pixels. The shutte is global, so the whole f ame is captu ed i a si gle simulta eous exposu e with o ow-by- ow timi g. The 3.0 mic o pixels a e la ge fo a 2MP se so , which helps collect light i dim wa ehouse aisles a d ight sce es. Togethe they make the module a AR0234 AGX O i Global Shutte Came a that delive s geomet ically faithful f ames to the platfo m. Teams that measu e object positio s o t ack motio betwee f ames be efit most f om that fidelity.
The o -boa d ISP is the seco d desig choice. It co ve ts the Baye data i to a fi ished image i side the module. Acco di g to the datasheet, it pe fo ms debaye i g, demosaici g, colo co ectio , co t ast co ectio , gamma co ectio , de oisi g, a d le s co ectio s, a d it suppo ts automatic fu ctio s such as auto exposu e a d auto white bala ce. The module outputs YUV422 at 8-bit o 10-bit depth. Because the image is al eady p ocessed, the came a is t eated as a YUV sou ce o Jetso . F ames each the captu e ha dwa e di ectly, a d the Jetso ISP is ot used. This saves the effo t of buildi g a d tu i g a ISP p ofile, a d it keeps colo behavio the same o a y host. The t ade-off is that tu i g is fixed i the module, so teams who wa t aw access a d custom ISP co t ol should choose a aw se so module, which Vadzo also offe s. Teams that a e u su e ca desc ibe thei pipeli e, a d Vadzo will ecomme d fi ished o aw output.
The module is a two-boa d solutio , with the came a module boa d a d a adapte boa d. The adapte comes i two va ia ts, o e fo 2-la e MIPI CSI-2 a d o e fo 4-la e. The 2-la e va ia t co ects di ectly to Raspbe y Pi a d NVIDIA Jetso developme t kits, a d the 4-la e va ia t p ovides the ba dwidth fo faste f ame ates. I 4-la e mode, the suppo ted video modes u at 120 f ames pe seco d at VGA, 60 at HD, a d 60 at Full HD. The came a co ecto o each ca ie boa d dete mi es which adapte fits, so teams should match the adapte to thei AGX O i came a, a d Vadzo suppo ts that selectio . Pla i g the co ecto choice ea ly also lets the e closu e a d cable outi g be desig ed o ce.
Co t ol a d e vi o me tal details complete the pictu e. Came a co t ols i clude b ight ess, exposu e, co t ast, sha p ess, satu atio , gamma, gai , white bala ce, de oisi g, a d LED flicke mitigatio , with a ti-flicke fo 50Hz a d 60Hz lighti g. Still images ca be captu ed as BMP files th ough a softwa e t igge . The module d aws betwee 0.80 W a d 1.45 W at 3.3 volts, ope ates f om mi us 40 deg ees C to 85 deg ees C, measu es 38 mm by 38 mm, is co ve tible to 32 mm by 32 mm, a d weighs 13 g ams without the le s. A S-Mou t (M12) holde accepts a 74 DFOV le s, a d two status LEDs show whethe the module is powe ed with o st eami g o actively st eami g. Humidity is ated at 20 to 80 pe ce t elative, o -co de si g.
It helps to be clea about what the datasheet states a d what it leaves to the i teg ato . The module delive s Full HD, HD, a d VGA video, a d the se so ‘s full 1920 x 1200 a ay is available as the maximum esolutio . The two status LEDs, the softwa e t igge fo still captu e, a d the BMP still fo mat a e pa t of the module, while the GPIO li es a e ot co ected by default. The le s is a fixed-focus S-Mou t (M12) assembly, so wo ki g dista ce should be chose du i g evaluatio . Teams that eed a diffe e t field of view ca eplace the le s with a othe M12 optic, a d Vadzo ca eview the choice agai st the se so .
