Enhancing Event-Based Vision for Logistics: Data Processing, Prediction Models Analisys and 3D Dataset Creation

Resumen

Event-based vision offers advantages in dynamic environments due to its asynchronous data acquisition and low-latency processing. This thesis presents an in-depth analysis of the Label-Efficient Object Detection (LEOD) algorithm, evaluating its performance on a custom logistics dataset. To ensure compatibility, LEOD has been modified, addressing inherited inefficiencies and optimizing its handling of event-based data in logistics applications.

A critical part of this work involves the automation of data preprocessing and annotation, streamlining dataset creation, label assignment, and event representation. These enhancements reduce manual intervention and improve reproducibility, while optimizations in data storage cut annotation space requirements by 90%, making large-scale dataset handling more efficient.

Experimental results confirm that LEOD meets expectations in semi-supervised learning and outperforms previous benchmarks in Weakly-Supervised Object Detection (WSOD), achieving higher detection precision than the Gen1 and 1Mpx datasets across all training sizes.

Additionally, this research develops a 3D dataset using Neural Radiance Fields (NeRF) technology, reconstructing environments from RGB data to extend event-based perception into three dimensions. Future research will focus on integrating LEOD with 3D event-based perception and improving segmentation techniques for enhanced detection precision.