AbstractThis thesis studies the impact of augmenting an abstract target detection model with a higher degree of realism on the ﬁdelity of the outcomes of camera network simulators in reﬂecting real-world results. The work is motivated by the identiﬁed trade-oﬀ between realistic but computationally expensive models and approximate but computationally cheap models. This trade-oﬀ opens the possibility for an al-ternative to augment abstract simulation tools with a higher degree of realism to capture both beneﬁts, low computational expense with a higher ﬁdelity of the out-comes.
For the task of target detection, we propose a novel decomposition method with an intermediate point of representation. This point is the core element of our model that decouples the architecture into two parts. Decoupling brings ﬂexibility and modularity into the design. This empowers practitioners to select the model’s fea-tures individually and independently to their requirements and camera settings. To investigate the ﬁdelity of our model’s outcomes, we build models of three detectors and apply on our lab-based image data set to create ground truth conﬁdences. By incorporating only a few more properties of realism, the ﬁdelity of our model’s out-comes improved signiﬁcantly when compared to the initial results in reﬂecting the ground truth conﬁdences.
Finally, to explore the implication of our high ﬁdelity target detection model, we select a case study from coverage redundancy in smart camera networks. High-lighting the performance of a coverage approach strongly relies on the reliability of target detection results. An underestimation in the performance of studied coverage approaches is determined by employing the standard abstract detection model when compared to the results of our model.
The identiﬁed underestimation in this study is one example of the general open concern in agent-based modelling about the unclear impact of simpliﬁed abstract models on the ability of the simulator to capture real-world behaviours.
|Date of Award||Sept 2020|
|Supervisor||Peter Lewis (Supervisor)|
- Target Detection
- Agent-based Modelling