Abstract
Crowd monitoring has been gaining attention among researchers for a wide range of applications. Such applications may include abnormal behavior detection for security reasons, monitoring of crowds for reporting, and people counting for structural and facility planning purposes. Traditional detecting models fail to detect people and count them in the crowded scenes. Deep learning networks have proven costly requiring energy and memory for computations beyond means in resource-constrained devices. Thus, we propose the YOLOv4 detection algorithm fused with Pruning technique during training and convolutional attention module. The proposed pruning strategy makes the model compact and light and thus reduces memory for the processing needs. Adding a Convolutional Block Attention Module to the YOLOv4 network improves detection accuracy by increasing the weights of valuable features while lowering the weights of invalid data. The Improved YOLOv4 then is not only used to locate everyone in the crowd, rather it counts them. The JHU-Crowd dataset of crowd images is used to train YOLOv4. The findings of this hybrid network are authenticated against the experimental results by the original YOLOv4. The hybrid YOLOv4 mAP is boosted by 21% and inference time is reduced by 11% and the loss reduced to 0.1. The proposed pruning mechanism reduces the number of computations and memory during training, which makes it better a better candidate for Real-Time monitoring.