Statistical modelling of algorithms for signal processing in systems based on environment perception (Dissertation)

Advan­ced dri­ver assis­tance sys­tems (ADAS), for exam­p­le auto­no­mous emer­gen­cy bra­ke (AEB) sys­tems, aim to sen­se a vehicle’s envi­ron­ment, under­stand the cur­rent traf­fic situa­ti­on and react appro­pria­te­ly in order to sup­port the human dri­ver. One cor­ner­stone for relia­ble sys­tems is an appro­pria­te hand­ling of uncer­tain­ties. Uncer­tain­ties ari­se for ins­tance due to noi­sy sen­sor mea­su­re­ments or the unknown future evo­lu­ti­on of a traf­fic situa­ti­on. This work’s objec­ti­ve is to con­tri­bu­te to the under­stan­ding of the­se uncertainties.

To this end, para­me­tric pro­ba­bi­li­ty dis­tri­bu­ti­ons are used to ana­ly­ti­cal­ly model and pro­pa­ga­te uncer­tain­ties at indi­vi­du­al parts of ADAS signal pro­ces­sing chains. Pre­vious works approach this most­ly with nume­ri­cal simu­la­ti­ons. An inher­ent draw­back is that only a con­cre­te imple­men­ta­ti­on of algo­rith­ms can be ana­ly­sed. Ana­ly­ti­cal model­ling on the other hand, as pur­sued in this the­sis, allows dra­wing more gene­ric con­clu­si­ons. One can for exam­p­le attempt to deri­ve upper per­for­mance bounds which app­ly to any imple­men­ta­ti­on of (sub-) opti­mal solutions.
This the­sis devi­ses pro­ba­bi­li­stic models of uncer­tain­ty in sel­ec­ted algo­rith­ms with a distinct role in cur­rent and future ADAS. The models are appli­ed to the deri­va­ti­on of sen­sor para­me­ter cons­traints for fea­ture-based loca­li­sa­ti­on for urban auto­ma­ted dri­ving and the ana­ly­sis of per­for­mance limi­ta­ti­ons in AEB systems.
First, envi­ron­ment per­cep­ti­on tasks are stu­di­ed. This com­pri­ses the detec­tion of obs­ta­cles with a ste­reo visi­on sen­sor and fea­ture-based loca­li­sa­ti­on. Ins­tead of a theo­re­ti­cal top-down ana­ly­sis, an envi­ron­ment sensor’s inac­cu­ra­cy can also be empi­ri­cal­ly eva­lu ated by com­pa­ri­son to a refe­rence sen­sor. Advan­ced off-line approa­ches can be lever­a­ged to obtain accu­ra­te refe­rence data from raw sen­sor mea­su­re­ments. The poten­ti­al of such methods, in con­trast to an on-line pro­ces­sing under real-time cons­traints, is stu­di­ed on the exam­p­le of a laser scan­ner sensor.
Second, the models of errors in indi­vi­du­al mea­su­re­ments are pro­pa­ga­ted to the uncer­tain­ty in esti­ma­tes of unob­ser­va­ble dyna­mic sta­tes, for ins­tance moti­on sta­te variables.
Third, the pre­dic­tion and risk assess­ment of traf­fic situa­tions is con­side­red. Novel expres­si­ons for per­for­mance bounds on the reco­gni­ti­on of seman­tic dri­ver inten­ti­ons for long-term moti­on pre­dic­tions are deri­ved. Moreo­ver, para­me­tric models of uncer­tain­ty in kine­ma­tic moti­on models for short-term pre­dic­tions are esti­ma­ted from empi­ri­cal dri­ving data. Based on the­se pre­dic­tions, the risk of an immi­nent col­li­si­on can be quan­ti­fied in terms of cri­ti­cal­i­ty mea­su­res. A com­pre­hen­si­ve approach to pro­ba­bi­li­stic model­ling of the­se risk metrics is deve­lo­ped. Per­for­mance bounds on the time­ly acti­va­ti­on of an AEB inter­ven­ti­on are then deri­ved by uncer­tain­ty pro­pa­ga­ti­on from per­cep­ti­on and sta­te esti­ma­ti­on to the cri­ti­cal­i­ty assessment.

Down­load Dis­ser­ta­ti­on (.pdf, 4,2 Mb)

Down­load Dis­pu­ta­ti­ons­vor­trag (.pdf, 1,3 Mb)


Beitrag veröffentlicht

in

von

Schlagwörter: