Design and implementation of a multimodal dataset processing pipeline for vehicle tracking model

Details

Supervisors

Macq, Benoît ; Craeye, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électricien, à finalité spécialisée

Abstract

With the rise of machine learning getting quality dataset to train new model on has always been the highest problem to tackle. This thesis proposes a solution to the pervasive challenge of procuring high-quality, labeled datasets for supervised machine learning. The solution takes the form of a multimodal data pipeline, capable of extracting labels from data derived from synchronized camera images and Doppler radar heatmaps. After delving into the fundamentals of machine learning, the thesis presents a critical review of current practices in image and signal processing, paying close attention to object identification techniques like YOLO for image processing and CFAR for signal processing. The thesis then introduces a unique data structure and elucidates the proposed data pipeline, providing a meticulous walkthrough of the process, including the stages of RGB image processing, radar heatmap processing, data merging, and sequential data analysis. It also suggest the incorporation of a transformer model to allow prediction capabilities. Experimental results are then evaluated and discussed, with a particular focus on the implications of dataset characteristics, established approximate ground truth annotations, and the necessary computational resources. The study concludes with a candid discussion of identified limitations and potential improvements, summarizing the study’s contributions to the domain multimodal data labelisation and preprocessing for supervised learning models.