ResNet50-YOLOv2-Convolutional Neural Network Based Hybrid Deep Structural Learning for Moving Vehicle Tracking under Occlusion

Authors

  • Latha Anuj, Dr. M. T Gopalakrishna

Abstract

The exponential rise in software computing and low-cost hardware technologies has broadened the horizon for the different applications serving human for reliable and efficient decisions. Amongst major contemporary applications, surveillance systems have demanded more efficient vehicle detection and tracking systems, especially under occlusion, which is common in contemporary dense urban traffic conditions. As a viable solution multiple cameras based target-tracking has gained widespread attention; though very less researches address target similarity, low-computation demands and time-efficiency, which is of utmost significance. Factually, identifying moving target region, extracting spatio-temporal features, learning, synchronizing results is always challenging. Though, deep learning methods have performed better for vehicle classification, its efficacy has remained unexplored, especially for moving target-tracking under occlusion-probability. However, learning deep features over different inputs and synchronization is tedious. Realizing such complexity, in this paper we developed a novel Hybrid Deep Structural Learning (HDSL) model by cascading layer-modified ResNet50 with YOLOv2 layer and Convolutional Neural Network (CNN) to perform moving vehicle tracking under occlusion. In our proposed method, ResNet50 functions as initial layer to extract multilayer features, while YOLOv2 operates for target identification, tracking and continuous labeling. Furthermore, to verify detected target and its tracking efficiency we applied CNN which distinguishes moving vehicles in its specific types. Noticeably, unlike classical YOLO layer based learning we designed YOLOv2 with normalized loss function based bounding box generation and labeling. ResNet5 and YOLOv2 helps achieving moving vehicle tracking under occlusion. MATLAB based simulation over standard benchmark datasets such as CityFlow, DETRAC MOCAT revealed that the proposed model achieves maximum tracking and online target classification (under occlusion) accuracy of 97.19%, while ensuring precision of 95.31%, recall 95.52% and F-Score of 95.41%.

Published

2020-10-16

Issue

Section

Articles