Discriminatively Trained Multi-source CNN Model for Multi-camera Based Vehicle Tracking Under Occlusion Conditions

Abstract

In this paper, a novel Discriminatively Trained Multi-Source CNN Model (DTM-CNN) is developed for multi-camera based vehicle tracking purpose. DTM-CNN performs pretraining of a gigantically large set of traffic videos to track ground truths for retaining region of interest (ROI) representation. Being a multi-source tracking method DTM-CNN embodies shared layers and multiple branches of source-specific layers to perform feature extraction and training. Here, source signifies each camera input with distinct training sequences, where each branch exhibits binary classification for ROI identification and tracking in each source. DTM-CNN trains each source input iteratively to achieve generic ROI representations in the shared layers. When performing tracking in a new sequence, DTM-CNN forms a new network by combining the shared layers with a new binary classification layer, which is updated online. It assists online tracking by retrieving the ROI windows arbitrarily sampled near the previous ROI state that enables DTM-CNN to exhibit continuous vehicle tracking even under short and long term occlusion.