From Impact to Insight : A Smart Vision-Based Projectile Hit Detection System for Bullet Impacts on Unmanned Surface Vessels

Sammanfattning

The increasing use of unmanned surface vessels (USVs) in maritime operations creates new needs for evaluating small caliber hard-kill countermeasure systems. Traditional impact detection methods are often limited by manual handling, low spatial resolution, or system complexity. This thesis presents a cost-effective, real-time camera-based system for detecting and localizing bullet impacts on a USV-type structure, intended as a feedback tool for training and evaluation. The system includes four main components: a custom target apparatus with a self-marking surface, a camera-based acquisition and transmission setup, a hybrid detection model, and a visualization interface. The target surface combines vinyl and acrylic paint to improve impact visibility and enable quick reset between tests. Video is captured using a Raspberry Pi edge unit and transmitted via 4G to a central processor, achieving near real-time operation. The impact detection model uses a hybrid method combining truncated convolutional neural network feature extraction with classical signal processing, balancing robustness and efficiency. During preliminary field testing of the prototype, the system achieved a peak F1-score of 0.84 on a constrained dataset. A user interface supports real-time feedback and post-session analysis through hit markers and heatmaps. Results indicate near real-time performance is achievable under stable field conditions, with end-to-end latency below two seconds. Overall, the work demonstrates the feasibility of a scalable vision-based approach for projectile impact detection on USV platforms and outlines future improvements such as edge processing, adaptive models, and more realistic maritime validation. The modular design also suggests potential applicability beyond the specific use case considered.