Performance Analysis of Object Detection and Tracking Algorithms for Traffic Surveillance Applications using Neural Networks

Abstract

The single object detection has been performed by using the concepts of convolution layers. A neural network consists of several different layers such as the input layer, at least one hidden layer, and an output layer. The dataset used for single object detection is the on-road vehicle dataset. This dataset consists of three classes of images which are Heavy, Auto and Light. The dataset consists of images of varying illuminations. The performance metrics has been calculated for the day dataset, evening dataset and night dataset. Multiple object detection has been performed using the You Only Look Once (YOLOv3) algorithm. This approach encompasses a single deep convolution neural network dividing the input into a cell grid and each cell predicts a boundary box and classifies object directly. The dataset used for multiple object detection is the KITTI dataset. It consists of 80 classes out of which five classes has been considered for this project which are: car, bus, truck, and motorcycle and train. Using the Multiple Object Detection concepts, tracking of vehicles was further implemented. The first frame of the video was taken and Multiple object detection was performed and in the further frames of the video the object was tracked using its centroid position. This has been developed using OpenCV and Python using YOLOv3 algorithm for the object detection phase.