|dc.description.abstract||Four adaptive metaheuristic optimization algorithms are proposed and demonstrated: Adaptive Parameter Particle Swarm Optimization (AP-PSO), Modified Artificial Bat (MAB), Differential Mutated Artificial Immune System (DM-AIS) and hybrid Particle Swarm Accelerated Artificial Immune System (PSO-AIS). The algorithms adapt their search parameters on the basis of the fitness of obtained solutions such that a good fitness value favors local search, while a poor fitness value favors global search. This efficient feedback of the solution quality, imparts excellent global and local search characteristic to the proposed algorithms.
The algorithms are tested on the challenging Articulated Human Tracking (AHT) problem whose objective is to infer human pose, expressed in terms of joint angles, from a continuous video stream. The Particle Filter (PF) algorithms, widely applied in generative model based AHT, suffer from the 'curse of dimensionality' and 'degeneracy' challenges. The four proposed algorithms show stable performance throughout the course of numerical experiments. DM-AIS performs best among the proposed algorithms followed in order by PSO-AIS, AP-PSO, and MBA in terms of Most Appropriate Pose (MAP) tracking error.
The MAP tracking error of the proposed algorithms is compared with four heuristic approaches: generic PF, Annealed Particle Filter (APF), Partitioned Sampled Annealed Particle Filter (PSAPF) and Hierarchical Particle Swarm Optimization (HPSO). They are found to outperform generic PF with a confidence level of 95%, PSAPF and HPSO with a confidence level of 85%. While DM-AIS and PSO-AIS outperform APF with a confidence level of 80%. Further, it is noted that the proposed algorithms outperform PSAPF and HPSO using a significantly lower number of function evaluations, 2500 versus 7200.
The proposed algorithms demonstrate reduced particle requirements, hence improving computational efficiency and helping to alleviate the 'curse of dimensionality'. The adaptive nature of the algorithms is found to guide the whole swarm towards the optimal solution by sharing information and exploring a wider solution space which resolves the 'degeneracy' challenge. Furthermore, the decentralized structure of the algorithms renders them insensitive to accumulation of error and allows them to recover from catastrophic failures due to loss of image data, sudden change in motion pattern or discrete instances of algorithmic failure. The performance enhancements demonstrated by the proposed algorithms, attributed to their balanced local and global search capabilities, makes real-time AHT applications feasible.
Finally, the utility of the proposed algorithms in low-dimensional system identification problems as well as high-dimensional AHT problems demonstrates their applicability in various problem domains.||en