AAI_2025_Capstone_Chronicles_Combined

17

Draw, Detect, Navigate ​

​

​

​

​

​

​

​

​

would likely be overcome in future iterations or workarounds, and would not impact

deployment to other mobile platforms.

The live performance testing demonstrated sufficient capability to enable more complex

data interactions such as route planning, adversarial simulations in games or strategic

contested environment planning, location design mock-ups, or foundations of understanding

human drawings where proximity of drawings encodes important relationships. Further work

would explore these emergent capabilities, comparisons of accuracy and speed of models on

live data, training with larger sets of images or more domain-specific imagery, and testing on a

broader range of platforms with lower compute resources available to further analyze viability

on edge devices. When turning future efforts toward more narrow domain applications where

the doodles present in the Quick, Draw! Dataset may not be sufficient, the problem of acquiring

sufficient volumes of training data would need to be addressed. Requesting leaders, creative

designers, mechanical engineers, or physicians to spend time hand drawing greater than 2,000

doodles of each needed subject not immediately known to a more general population could be

prohibitively costly and time consuming. The use of an additional model to generate individual

synthetic doodles to include in the synthetic multi-doodle image generator could be utilized to

address this potential limitation.

This project aims at real-time detection and classification of doodles with the purpose of

determining optimal routes from one point to another. This means route optimization directly

depends on the high performance of detection and classification. Consequently, poor model

performance could severely impact both the accuracy and speed of generating optimal routes.

For instance, low accuracy or frequent misclassification might lead to suboptimal routing

decisions or delays, which could have severe consequences in emergency scenarios where

prompt, accurate decisions are critical to saving lives. Moreover, the reliability and robustness

of the model in varying real-world conditions, such as differences in lighting, drawing styles,

image quality, or unexpected doodle variations, will significantly influence practical outcomes.

False positives could result in unnecessary diversions, while false negatives could mean failing

44