The purpose of wide-area tracking problem is to track pedestrians that appear on cameras that overlap or do not overlap, regardless of the time interval or person density.
In a single camera tracking, data association using overlapping of the detectio...
The purpose of wide-area tracking problem is to track pedestrians that appear on cameras that overlap or do not overlap, regardless of the time interval or person density.
In a single camera tracking, data association using overlapping of the detection boxes is used to solve the tracking problem, but still has appearance ambiguity issues.
However, wide-area tracking requires a tracking scheme that focuses on the appearance similarity of humans, without the use of overlapping of detection boxes.
In this dissertation, we propose the tracking scheme for the Wide-area Multi-Pedestrian Tracking (WaMuPeT).
To achieve the WaMuPeT, we propose the trajectory matching in overlapping camera settings (Ch. 3), non-overlapping camera settings (Ch. 4) and robust trajectory matching in dense scene settings (Ch. 5).
In trajectory matching in overlapping camera settings (Ch. 3), we propose a novel deep-learning architecture for accurate 3-D localization and tracking of a pedestrian using multiple cameras.
The deep-learning network is composed of two networks: detection network and localization network.
The detection network yields the pedestrian detections and the localization network estimates the ground position of a pedestrian within its detection box.
In addition, an attentional pass filter is introduced to effectively connect the two networks.
Using the detection proposals and their 2-D grounding positions obtained from the two networks, multi-camera multi-target 3-D localization and tracking algorithm is developed through min-cost network flow approach.
In the experiments, it is shown that the proposed method improves the performance of 3-D localization and tracking.
In trajectory matching in non-overlapping camera settings (Ch. 4), we propose a novel re-ranking method using a ranking-reflected metric to measure the similarity between two ordered sets of $K$-nearest neighbors (OKNN).
The proposed metric for ranking-reflected similarity (RSS) reflects the ranking of the shared elements between the two OKNNs.
Using RSS, a re-ranking procedure is proposed that prioritizes galleries having neighbors similar to a probe's neighbor in the perspective of ranking order.
In the experiment, we show that the proposed method improves the Re-ID accuracy by add-on to the state-of-the-art methods.
In robust trajectory matching in dense scene settings (Ch. 5), we propose a novel framework for multi-pedestrian tracking to generate robust trajectories in dense scene.
In the proposed tracking method, we propose the tracking method based on the trajectory matching by the strategy of divide and conquer method.
In this strategy, short-term, mid-term and long-term trajectories are generated by each trajectory merging stages, respectively.
Also we propose a novel deep-feature matching method called stable boundary selection (SBS).
In SBS matching, the detections are clustered by the group similarity of deep features, so that robust trajectories can be generated.
With the smoothing algorithms and the detection restoration algorithm, the proposed tracking method shows the state-of-the-art tracking accuracy in three public tracking dataset.