Pruning Based Improved Pedestrian Detection in Video Surveillance

Real-time object identification applications are using methods for image processing more and more as a result of advancements in computer technology. Results for pedestrian identification were improved by the use of deep learning techniques. To locate human beings quickly and precisely, this study suggests employing pruning to increase the Yolo models' efficiency in pedestrian recognition. This method would allow the device to function seamlessly when used for video surveillance. This method uses one stage detector, YOLOv5, to better suit the task situations and adjust to the real-time demands. To improve the possible model sparsity, 1 × 1 kernel weights are pruned using a pruning approach. The findings show that, following pruning, this model attains a 30% sparsity in this model, which indicates that 30% of the weight parameters in the nn. Conv2d layers of the algorithm are equal to 0. Inference time is nearly constant and soft-nms is applied for better ROI results. In addition, a thorough result and discussion are given to demonstrate the efficacy of this methodology regarding mAP, latency, energy consumption, and attained sparsity.