Research Assistant, Institute of Information Science (IIS), Academia Sinica
Bio-Industrial Mechatronics (BIME), National Taiwan University
PERSON RE-IDENTIFICATION FOR PEDESTRIAN TRACKING
Institute of Information Science, Academia Sinica
PROJECT IN PROGRESS:
DEEP-BASED PEDESTRIAN RE-IDENTIFICATION & TRACKING SYSTEM ON A DISTRIBUTED NETWORK
Multi-target multi-camera tracking (MTMCT) has a wide variety of applications, for instance, surveillance, sport player tracking, and so on.
In MTMCT tasks, same individuals are expected to be identified and tracked across distinct camera views and over occlusions (person re-identification).
Occlusion and its consequential effects is the most challenging part. Occlusion of individuals fragments the tracks and may result in re-identification errors.
Even though sometimes individuals are never totally occluded, track fragmentation and re-identification errors still occur. My current research aims for solutions to partial occlusion problems.
BACKGROUND
State-of-the-art pedestrian re-identification methods are based on detections.Â
However, partial occluded pedestrian detections could bring unfavorable results to the re-identification results when "appearance vectors", a.k.a. "embedded features", are extracted using a deep neural network model.
Examples of "partially occluded" individuals and possible embedding results are illustrated in the image below.
PRELIMINARY IDEA
Therefore, we come up with the following framework, as illustrated in the diagram.
Taking advantage of pose estimators, we could be able to evaluate visible parts in each detection, and then modify embedding and metric comparison methods, creating a more specific and accurate appearance metric.
EXPERIMENTAL SETTINGS
Evaluation on dukeMTMC-ReID dataset (Pedestrian detections & ground truth available).
Detector: YOLO V3
Pose estimator: AlphaPose (RMPE, detection-based pose estimator)
​
Step (1): Examine our solution towards partial-occlusion yields better performance.
Step (2): Train network based on part-visibility & partial embedding ideas.
Pose estimation based on human detections.
"False positive" pose estimations.
Occluded identities may also be undetected.
Pose estimation based on human detections.
CASEÂ DEMONSTRATIONS
To validate & demonstrate our preliminary ideas, we tested our framework on several example images of people with different poses.
Back cases: Only ears are detectable. Thus we can regard the head part as either visible or not (by adjusting visibility score thresholds). Other parts are determined as expected.
Occluded cases: After modifying the yolo detector, partial occluded cases are detected as we expected.
As for the pose estimator, the invisible parts score a very low probability and are regarded "invisible".
Other pose cases: Though only keypoints on single side are detectable, there is no effect on visibility determination due to the average threshold strategy.
APPEARANCE EMBEDDER
Objective of the appearance metric is to only embed and compare visible parts of individuals.
According to proposed concept of AlignedReId, we define "body parts" as a set of horizontal strips.
Visible parts are aligned and resized according to pose hints, while occluded/invisible parts are disregarded.
FRAMEWORK CONSTRUCTION & EVALUATION
February 11 ~ Present
Currently in progress & updates actively!