Improving Efficiency of DNN-based Relocalization Module for Autonomous Driving with Server-side Computing

Abstract: In this work, we present a novel framework for camera relocation in autonomous vehicles, leveraging deep neural networks (DNN). While existing literature offers various DNN-based camera relocation methods, their deployment is hindered by their high computational demands during inference. In contrast, our approach addresses this challenge through edge cloud collaboration. Specifically, we strategically offload certain modules of the neural network to the server and evaluate the inference time of data frames under different network segmentation schemes to guide our offloading decisions. Our findings highlight the vital role of server-side offloading in DNN-based camera relocation for autonomous vehicles, and we also discuss the results of data fusion. Finally, we validate the effectiveness of our proposed framework through experimental evaluation.

• The paper proposes an offloading framework for DNN relocalization, reducing onboard computational load by transferring intensive tasks to server-side computing.
• The paper demonstrates through experiments that MapNet-based systems achieve increased inference speeds and timely pose estimations on low-power devices.
• The paper highlights potential future extensions to other SLAM modules and cross-device inferencing, advancing cloud-edge collaboration in autonomous driving.

Introduction to DNN-based Relocalization

In the domain of autonomous driving, highly precise environmental mapping and accurate vehicle self-positioning are critical for safety and efficiency. Traditional global navigation satellite systems (GNSS) like GPS can sometimes be unreliable due to environmental interferences, leading to the adoption of visual SLAM (Simultaneous Localization and Mapping) technology for better accuracy. The relocalization module of SLAM systems is key in determining the vehicle's position on a high-precision map. However, the computational intensity of current DNN-based relocalization methods can be prohibitive for deployment on small devices with limited processing power and energy budgets, such as those used in autonomous vehicles.

Efficiency Through Edge Computing

This paper introduces a novel framework to improve the efficiency of DNN-based camera relocalization for autonomous vehicles using server-side computing, often referred to as edge cloud collaboration. This approach strategically offloads portions of the neural network computation to the server. Different offloading schemes are evaluated based on inference times, which helps optimize the distribution of computational tasks between the vehicle and the server. This server-side processing reduces the computational load on the vehicle, thus leading to improvements in both the speed of processing frames and the volume of pose information received by the vehicle within the same period. Experimental evaluations provide validation for the proposed framework.

Experimental Methodology

Through simulation experiments that engage both a low-power mobile device and a more capable computing server, the framework's offloading strategy is put to the test. Utilizing commonly recognized datasets from the autonomous driving domain, the proposed offloading framework demonstrates an ability to considerably increase the operational speed of the MapNet-based camera relocalization system without compromising relocation accuracy. The experiments suggest that using server-side computing to offload the computation-intensive reasoning process allows autonomous vehicles to benefit from rapid pose estimations, thus improving navigation and driving decisions.

Impact and Future Work

The study concludes that offloading DNN-based camera relocalization computation to a server significantly enhances reasoning efficiency and could have a wide-ranging impact on fields such as cloud robotics. The enhanced reasoning frequency and extended coverage capabilities contribute to the approach's practical applications. Additionally, data fusion using GPS and network outputs presents a promising avenue for improving the auxiliary positioning information of autonomous vehicles. As future work, the paper advocates for extending this offloading strategy to other SLAM system modules and further exploring cross-device model inference, shedding light on the importance of cloud-edge collaboration in advancing autonomous vehicle technologies.