Tracking

Multi-Target Multi-Camera Tracking (MTMCT) with Person Re-Identification (Re-ID)

We have implemented the Features for Multi-Target Multi-Camera Tracking and Re-Identification(Re-ID) paper to achieve this functionality. This system aims to determine position of every person in every frame on all times. This system provided us the benefit of Re-ID over traditional optical flow algorithm. Final result of the algorithm is all the detected trajectories rendered into a single frame, where same color trajectories correspond to same person.

This algorithm learns good features from both MTMCT and Re-ID with the help of a Convolution Neural Network. MTMCT classifies a pair of images being identical or not while Re-ID ranks distances to query i.e. the distances between a and a feature which is identical to it should be smaller than the smallest distance between a and a feature which is not identical to it. This method performed better in comparison to the other approaches because it makes use of weighted triple loss for training, which is more stable and accurate than fixed weight variants.

DATA SET: Duke MTMC

It is a large-scale tracking dataset recorded on the Duke University campus featuring 2.8k identities, of which 1.8k belong to the training/validation set. The dataset was recorded by 8 cameras with 1080p 60fps image quality and the evaluation is done on disjoint fields of view.

APPROACH

Optical flow and its variants

To begin with, the conventional tracking algorithms including optical flow and its variant (dense optical flow) were tested for performance with single camera setup. While they enabled point tracking, detecting and tracking points pertaining to the same person was the major challenge faced.

Target scene

Dense Optical flow based person tracking

Optical flow based person tracking

Deep Learning based detection and tracking

Having performed a comprehensive literature survey, we could conclude that simultaneous detection and tracking could be best performed using deep learning based algorithms. Deep tracking is currently a widely researched problem statement. Two popular data-sets pertaining to multi-camera tracking were identified, namely, EPFL data set and Duke MTMC dataset.

The Deep Occlusion based implementation on EPFL Dataset performs people detection using Probabilistic Occupancy Map(POM). The trajectory is generated through K-Shortest Path (KSP). This approach was found to incur large resource complexity.

To this end, DeepCC in the paper found to be a good alternative. The key challenges addressed by this approach are:

Multi-target multi-camera tracking (MTMCT) :
1. problem of determining who is where at all times
2. Required to analyse data for strategic planning

2. Person Re-identification (Re-ID) - given a query image, localize the person in images captured by the multiple cameras.

Clearly, the algorithm not only enables us to detect and track people across frames but also helps perform person re-identification. This could be a crucial addendum for our end application which aims to aid data-analytics in profitable product placement in super-markets.

Results for person detection using Deep Occlusion

Result and Analysis for DeepCC

Trajectories across all cameras

Camera 1: Trajectories

Camera 2: trajectories

Google Sites

Report abuse