Parameter Optimization for Partitioning Algorithms in Object Detection

Fast object detection is essential for safety-critical applications. Effective partitioning of video frames captured by cameras acting as sensors can enhance the performance of object or target detection in a parallel processing system, ensuring that no node becomes a bottleneck. More specifically, algorithms that utilize frame-specific information, such as entropy, to estimate the workload of frames and then partition them accordingly perform better than those that partition frames randomly. The objective of this paper is to optimize partitioning algorithms for big data applications to achieve the best performance. Parameters, such as neighborhood size for Entropy-Based partitioning, are determined experimentally to minimize overheads while ensuring accurate workload estimation. A novel hybrid partitioning algorithm is proposed to enhance the performance of previously proposed frame-specific partitioning algorithms. Workload of the frames are estimated using object detection models. To identify the most effective object detection model, pretrained models such as You Only Look Once (YOLO) and Vision Transformer (ViT) are deployed within the Apache Spark parallel processing framework and the model with the best processing time and accuracy is chosen for workload estimation. To demonstrate the robustness of the proposed techniques, the partitioning and frame removal algorithms are evaluated under noisy frame conditions and in a Graphics Processing Unit-accelerated environment. The proposed partitioning and frame removal algorithms performed well in noisy frame conditions where workload estimation was more difficult and achieved approximately 50 % better performance when a GPU was utilized.