Detection-based multi-object tracking in presence of unreliable appearance features

Multi-object tracking (MOT) is the task of estimating the trajectory of several objects as they move around a scene. MOT has gained in interest due to its potential in many disciplines such as surveillance, sport analysis, human computer interface, biology, etc. This thesis considers MOT in a scene captured by one or several sensors. It assumes that prior detections of the targets are available, and a set of features characterizing the appearance of the detections, have been extracted. In contrast to previous related works, we aim at formalizing the scenarios in which the reliability or even the availability of such appearance features vary over time. Our contributions, briefly explained below, are all related to the exploitation of sporadic and noisy features in a graph-based framework. Our first major contribution proposes an iterative hypothesis testing (IHT) framework that embeds shortest-path computations into a hypothesis testing procedure. Each hypothesis assumes that the appearance of the target is de- fined by that of a selected node, called key-node. Given this assumption, the cost of going through a node that is different (similar) from the key-node is increased (decreased) to favor the selection of a path that is consistent with the appearance of the key-node. The shortest-path is validated only when it is sufficiently better than any alternative path. Doing so, we progressively aggregate the detections into tracklets while taking advantage of their features, even when they are sporadic and/or affected by non-stationary noise. In the subsequent target recognition, we utilize the unreliability of the appearance features to prioritize the message passing while assigning reliable identities to the tracklets. As a second contribution, we propose a discriminative label propagation (DLP) framework to propagate labels across the detections in a way that is consistent with a number of complementary graphs that captures the various relationships, e.g., similarities or dissimilarities between the detections in terms of space, time and/or appearance between the detections. The resulting cost function is a difference of convex functions, which is efficiently solved using majorization-minimization techniques. We propose to decompose this global objective function into node-wise sub-problems. This not only allows a computationally efficient solution, but also supports an incremental and scalable construction of the graph, thereby making the framework applicable to large graphs and practical tracking scenarios. Moreover, it opens the possibility of parallel implementation.(FSA - Sciences de l'ingénieur) -- UCL, 201