Exploring millimeter-wave radar algorithms for obstacle detection and tracking

Rafael Pedro Carvalho

Key information

Authors:

Rafael Pedro Carvalho (Rafael Pedro Carvalho)

Supervisors:

Alberto Manuel Martinho Vale (Alberto Manuel Martinho Vale)

Published in

June 27, 2024

Abstract

A detecção e o seguimento de obstáculos em tempo real continuam a ser desafios importantes na indústria automóvel. Tradicionalmente, sensores como os LiDARs e as câmaras dominaram este espaço, no entanto, avanços tecnológicos recentes despertaram um novo interesse na tecnologia de radar, mais especificamente nos sensores de radar millimeter-wave (mmWave). Os sensores de radar mmWave mostram ser mais robustos a condições climatéricas adversas que outros sensores tipicamente usados. Este trabalho foca-se no desenvolvimento de um algoritmo de detecção e seguimento de obstáculos que usa somente dados provenientes destes radares. O algoritmo desenvolvido acumula alguns frames recentes de dados destes radares, de seguida usa DBSCAN para a detecção de obstáculos e Filtros de Kalman para o seguimento. Os resultados obtidos mostram que esta abordagem é eficaz a seguir diferentes tipos de obstáculos. No entanto, mostra existir uma maior capacidade em seguir obstáculos que se afastam do sensor comparativamente a obstáculos que se aproximam do sensor. Esta discrepância é causada por algumas características intrínsecas da tecnologia dos radares. Ainda assim, com uma configuração adequada do radar e com algum processamento dos dados, esta diferença pode ser atenuada. Esta abordagem serve como uma maneira de avaliar as capacidades dos radares mmWave em sistemas de detecção e seguimento de obstáculos enquanto Redes Neuronais robustas treinadas para este problema em específico ainda não estão prontas, principalmente devido à grande quantidade de dados necessários para um treino eficaz. Robust real-time obstacle detection and tracking remain significant challenges in the automotive industry. Historically, sensors like LiDAR and cameras have dominated this space, but recent technological advancements have enabled the resurgence of radar technology, specifically millimeter-wave (mmWave) radar sensors. MmWave radar sensors prove to be more robust to adverse weather conditions than other commonly used sensors. This work focuses on developing an obstacle detection and tracking algorithm that relies exclusively on data from these radars. The developed algorithm starts by accumulating recent frames of radar data and then employs DBSCAN for obstacle detection and Kalman Filters for tracking. Results indicate that this approach effectively tracks various types of obstacles. However, it demonstrates superior performance in tracking obstacles moving away from the sensor compared to those moving toward it. This discrepancy arises from the intrinsic properties of radar technology. Nonetheless, with proper radar configuration and processing, this difference can be mitigated. This approach serves as a method for evaluating the capabilities of mmWave radar systems in obstacle detection and tracking while robust Neural Networks trained specifically for this problem are not yet feasible, mainly due to the extensive data requirements for proper training. MmWave radar sensors prove to be more robust to adverse weather conditions than other commonly used sensors. This work focuses on developing an obstacle detection and tracking algorithm that relies exclusively on data from these radars. The developed algorithm starts by accumulating recent frames of radar data and then employs DBSCAN for obstacle detection and Kalman Filters for tracking. Results indicate that this approach effectively tracks various types of obstacles. However, it demonstrates superior performance in tracking obstacles moving away from the sensor compared to those moving toward it. This discrepancy arises from the intrinsic properties of radar technology. Nonetheless, with proper radar configuration and processing, this difference can be mitigated. This approach serves as a method for evaluating the capabilities of mmWave radar systems in obstacle detection and tracking while robust Neural Networks trained specifically for this problem are not yet feasible, mainly due to the extensive data requirements for proper training.