AHP-PSO: Adaptive hybrid particle swarm optimization for real-time cross-domain object detection
Abstract
Real-time object detection is critical in smart systems like self-driving cars, robots, and security. Conventional deep learning architectures like YOLO, SSD, and Faster R-CNN provide high accuracy in well-controlled scenarios but are computationally intensive and don't learn well under dynamic or resource-constrained conditions. These deep learning methods tend to be based on large annotated datasets and high-end hardware, preventing their use in real-time, edge, or resource-constrained settings. Though Particle Swarm Optimization (PSO) is a lightweight alternative as it is simple and capable of a global search, traditional PSO is plagued with premature convergence, the absence of temporal consistency, and non-competence on high-dimensional visual tasks. Current versions, such as QPSO and SPSO, address these problems partially but are not generally adaptable across domains. This work introduces AHP-PSO (Adaptive Hybrid Particle Swarm Optimizer), a new paradigm that unites quantum-behaved particle dynamics, motion-aware updates, and fitness-weighted inertia modulation. Its major contribution is its adaptive per-particle inertia mechanism, feedback-coupled swarm expansion, and domain-sensitive fitness modeling, which result in resilience and efficiency under diverse visual conditions. As compared to the state-of-the-art techniques, AHP-PSO is more accurate by 12%, has improved IoU by 9%, and converges 18% faster. With 0.80 F1-score and 15 FPS, it maintains a good trade-off between speed and accuracy, indicating scalability and robustness across aerial, underwater, and road environments