Crowd counting using deep learning based head detection

Abstract

Accurate and fastest object detection models are in high demand due to its wide variety of applications in the fields of computer vision, such as pedestrian detection, video surveillance and especially crowd counting applications. Automated crowd has been and continues to be a difficult problem for autonomous visual surveillance for many years. In the relevant literature, a substantial amount of research has been undertaken on the subject of crowd-counting and different architectures have been proposed for accurate and timely detection of heads in a crowd. Most of the approaches are based on regression, segmentation, image processing, machine learning techniques, counters and sensor-based models. Although the advancements in infrastructure has significantly improved the prediction accuracy but small heads are often missed by most of the proposed architectures in the literature. Scale invariance and high miss detection rates for small objects leads to the inaccurate results. The purpose of this research is to provide an accurate and fastest detection model for crowd counting by focusing on human head detection in real time scenarios acquired from publicly available datasets of Casablanca, Hollywood-Heads and Scuthead. In this study, we have tuned a yolov5 which is a deep convolutional neural networks (CNN) based object detection architecture by improving the mAP, precision and recall. The loss factors are reduced and accurate results are achieved by accurate tuning of hyper-parameters. Transfer learning approach is used for fine-tuning the architecture. From the experimental results, it can be seen that this yolov5 architecture showed significant improvements in small head detections in crowded scenes as compared to the other baseline approaches such as that Faster R-CNN and VGG-16 based SSD MultiBox Detector. In Faster R-CNN, features are extracted in the last layer therefore image resolution is decreased and small objects are not detected while yolov5 perform slicing of feature maps in the backbone region Therefore, small heads are detected accurately. Another main contribution of our research is use of merge dataset which include every kind of heads that is medium, large and small