Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Object Identification: A Bayesian Analysis with Application to Traffic Surveillance By Timothy Huang and Stuart Russell University of California at Berkeley 1 • Object Identification: A and B are the same object? • Object Recognition: A ? Class 1 ? Class 2 ? …..... Class k • Is there difference in practice? 2 • k moving objects: trajectory of object i modeled by r.v. Prh h h1 , , hk is a prior on object trajectories • Agent makes observations: oi hi o o1 , ,on is observation of some object 3 • Given o1 , ,oa , ,ob ,on • Interested in: Pr a b ,o Pr a b o Pr o 1 Pr a b , o h Pr h dh Pr o 4 Application to traffic surveillance Camera D Camera U o o , , o u u 1 u n o o , , o d o o ,o u d d 1 d n Find: average travel time, origin/destination counts 5 • Instead of modeling trajectories: u 1 o d 1 o u 2 o d 2 o u n o d n o • Matching observations is less general than modeling trajectories 6 S is the set of all possible matchings s 1,i ,2, j ,,n,l S Pr s is a uniform prior on S 7 Pr a b o , o u 1 Pr o u , o d 1 Pr o u , o d d Pr a b ,o u ,o d u d Pr o , o u d Pr a b , o , o ,s sS u d Pr o , o s Pr s sS : a ,b s 1 d u u Pr o o , s Pr o s Pr s Pr o sS :a ,b s d u Pr o o ,s sS : a ,b s 8 Assumption: Pro Pr o o , s d u d j o ,i j u i i , j s Pr a b o ,o d u d u Pr o j oi ,i j sS : a ,b s i , j s Computationally expensive: (n-1)! matchings to consider 9 t , time of observation x , lane position w , vehicle width l , length height d oj h , mean color hue c , mean color saturation v , mean color value C , histogram of color distribution d j d j d j d j d j d j d j d j For each feature, Pr x x ,i j d j u i 10 In particular: Pr t t , x , x ,i j d j u i u i d j N xu ,x d , x u ,x d t t i xiu ,x dj xiu ,x dj j i j d j u i Standard deviation and mean of predicted link travelu times starting upstream in lane x i d ending downstream in lane x j 11 Appearance probability: Pr o o ,i j d j u i Pr x x ,i j Pr C C ,i j d j u i d j u i 12 System learns parameters online: Downstream Upstream Lane 1 Lane2 Lane 1 Lane2 2 Pr x 1 x 1,i j 6 3 d u Pr x j 2 xi 2 ,i j 4 d j u i 13 Exponential forgetting: lane x u i matches lane x d j time t xiu , x dj xiu , x dj 1 t controls how fast we forget 14 Matching • Aim: find pairs (a,b) s.t. Pra b o 1 • Formula computed in previously computationally intractable • Can find most probable complete matching in n 3 time by weighted bipartite matching • In best matching, Pr a b o is not necessarily high for all (a,b) 15 “Leave one out” heuristic: a,b,c • Upstream observations: Downstream observations: x , y , z • Best assignment: x a, y b, z c with probability p • Forbid match x a and compute new best assignment x b, y a, z c with probability p • If p p threshold accept match (a,b) • Repeat for all matched pairs 16 m upstream and n downstream observations • n dummy upstream and m dummy downstream observations • odj , dummy vehicle is “new”, i.e. entered below upstream camera • dummy,o u i vehicle has left before downstream camera 17 • Pr dummy o u , probability to exit j • Pr oid dummy Pr oid , where is some coefficient and Pr oid is prior appearance probability 18 Results • With on-ramps and off-ramps 14% matched ---- 100% accuracy 80% matched ---- 50% accuracy • Without on-ramps and off-ramps 37% matched ---- 100% accuracy 80% matched ---- 64% accuracy • Link travel time --- accurate within 1% for 2 mile distance --- no bias based on speed 19