Low-cost multi-camera object matching

Abstract

We propose an object matching approach aimed at smartphone cameras that exploits the well-known concept of local sets of features for object representation. We also enable the temporal alignment of cameras by exploiting the frames of detected objects to group objects appeared in the same time interval for the assignment within each camera. The proposed approach does not need training thus making it suitable for matching during short temporal intervals. We use both outdoor and indoor datasets for the evaluation, and show that the proposed method reduces up to 95% the amount of information to be stored and communicated.

Re-identification with multiple source-cameras

“…We use two publicly available person datasets: the WARD dataset [13] and the Torch dataset [12]. WARD contains 70 persons from three non-overlapping fixed-cameras with the (e) (f) Fig.…”

“…3. Ranked ROC curves for re-identification in three-camera settings using the existing approaches: PRDC [3], ASFI [26], rankSVM [4] and CoPE [10] compared with the proposed association approach using, (a-c) WARD dataset [13] (|Pn| = 30), and (d-f) Torch dataset [12] (|Pn| = 24). Each camera detects persons that come from the other two as source-cameras such that persons appear in (a, d) C 1 from C 2 and C 3 , (b, e) C 2 from C 1 and C 3 , and (c, f) C 3 from C 1 and C 2 .…”

“…Ranked ROC curves for re-identification in three-camera settings with 25% additional persons (FPs) using the existing approaches: PRDC [3], ASFI [26], rankSVM [4] and CoPE [10] compared with the proposed association approach on (a) WARD [13] and (b) Torch dataset [12]. increases (Fig.…”

“…Features include color, texture and shape of an object [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12]. Features to be used for re-identification need to present a certain level of invariance to account for the pose, illumination and size variations that are likely to occur in different camera views.…”

“…Finally, we select the correct match from the obtained likelihoods L n by optimal assignment using the Hungarian algorithm [12]. The algorithm performs group assignment without repetition such that the summation of likelihoods between all the assigned pairs in a group remains minimum.…”