Online Learned Discriminative Part-Based Appearance Models for Multi-human Tracking

Abstract

Abstract. We introduce an online learning approach to produce discriminative part-based appearance models (DPAMs) for tracking multiple humans in real scenes by incorporating association based and category free tracking methods. Detection responses are gradually associated into tracklets in multiple levels to produce final tracks. Unlike most previous multi-target tracking approaches which do not explicitly consider occlusions in appearance modeling, we introduce a part based model that explicitly finds unoccluded parts by occlusion reasoning in each frame, so that occluded parts are removed in appearance modeling. Then DPAMs for each tracklet is online learned to distinguish a tracklet with others as well as the background, and is further used in a conservative category free tracking approach to partially overcome the missed detection problem as well as to reduce difficulties in tracklet associations under long gaps. We evaluate our approach on three public data sets, and show significant improvements compared with state-of-art methods.

Multi-object Tracking via Constrained Sequential Labeling

“…Existing multi-object trackers typically use points, tracklets, or object detections [21,17,12,2,19]. The previous section highlighted the challenges of using object detections with significant occlusion and pose variation.…”

“…1 shows, this can help disambiguate two people partially occluding each other. These mid-level features, however, present fundamental challenges to the common network-flow, and related formulations of data association which use object detections (e.g., [14,13,2,19,21,17] ). In particular, a person to be tracked is typically represented by an unknown number of mid-level features, which split and merge in both space and time.…”

Multi-pose multi-target tracking for activity understanding

“…Traditional association approaches have used generic appearance-based features, such as color histograms, local texture histograms or part appearance models as a means of defining a generic distance metric. However, recent work has shown that discriminative object-specific descriptors can be beneficial for robust tracking [10,12,14,22]. Following the direction of previous work, we implement an additional data association step that merges shorter tracklets over long temporal duration by leveraging invariant appearance information.…”

“…Advances in robust category-specific object detectors [5,6] have motivated the tracking-by-detection paradigm, where robust detectors can act as strong observation models in tracking frameworks. In particular, recent work has shown that a single coarse part-based model (e.g., 5 to 15 parts) [7,10,22] is well-suited for detecting, representing and tracking upright people. While these approaches are effective for urban scenarios, such as pedestrians walking on sidewalks or people in subway stations, difficulties arise when people perform other activities like riding a bike, digging a hole, or pushing a cart.…”

Partial Occlusion Handling for Visual Tracking via Robust Part Matching

“…Among them, methods part-wisely modeling object appearance [1,13,9,27,18,29,25] become more popular partially because of their favorable property of robustness against partial occlusion. Indeed, when there exists partial occlusion, some parts of the object remain visible which provide reliable cues for tracking.…”