Hybrid Residual-Deep-Convolutional Neural Network for Multiclass Vehicle Detection, Tracking and Classification: Resyalexnet

Abstract

In this paper, a highly robust hybrid residual-deep convolutional neural network (CNN) based vehicle detection, tracking and classification system is developed to enable end-to-end traffic surveillance purposes. Architecturally, our proposed model encompasses ResNet50, YOLOv2 and AlexNet-CNN, implemented in sequence to perform vehicle detection, tracking and classification. Considering operating environment complexity such as illumination difference, clutter, and occlusion, ResYAlexNet model at first applied ResNet50 as residual network which extracts fine-grained and detailed information from input frames consecutively. Subsequently, unlike classical deep learning methods where authors directly performed classification over the extracted features, we applied adaptive bounding box and loss-function enhancement based YOLOv2 model which helped identifying region of interest (ROI), target labelling, and its continuous target tracking. This approach helped performing optimal ROI tracking even under occlusion condition. Once generating the bounding boxes for the vehicles in each frame, we applied a well-known transferable deep learning method named AlexNet-CNN with five convolutional and three fully connected layers for learning and target classification. In the proposed model we applied AlexNet-CNN onto the target bounding box to avoid major background conditions and allied complexities. Additionally, it helped in achieving better ROI training and fast target (i.e., vehicle) tracking and classification. MATLAB 2019b based simulation over the different traffic videos with different vehicle densities and lighting conditions, exhibited higher accuracy (95.61%), precision (0.9722), recall (0.9683) and F-measure (0.9672), signifying its robustness towards real-time traffic surveillance purposes.