Object Detection Related to Construction Activity using Deep Learning

Abstract

Object detection problem has seen lot of applications recently with growing requirement to make detection real time. Present work focuses upon detection of objects related to construction activity that may be encountered unexpectedly on roads and cause traffic disruption. Recent developments in deep learning networks have made detection of complicated objects feasible within timing constraints suitable for Real time applications. This work uses Yolo algorithm in the form of darknet architecture to perform detection on a custom built dataset of traffic signs and cones. Yolo as an object detection algorithm has proved its applicability to real time tasks. The train and test data sets were compiled using standard open source data sets with labels in the form of bounding boxes around interesting objects. This work also tells importance of Yolo in video based object detection and highlights its limitations. The training and testing

process resulted in average loss of 0.1010 with mAP stabilizing at 91.21%