Investigating the Effect of Sensor Modalities in Multi-Sensor Detection-Prediction Models

01/09/2021
by   Abhishek Mohta, et al.
0

Detection of surrounding objects and their motion prediction are critical components of a self-driving system. Recently proposed models that jointly address these tasks rely on a number of sensors to achieve state-of-the-art performance. However, this increases system complexity and may result in a brittle model that overfits to any single sensor modality while ignoring others, leading to reduced generalization. We focus on this important problem and analyze the contribution of sensor modalities towards the model performance. In addition, we investigate the use of sensor dropout to mitigate the above-mentioned issues, leading to a more robust, better-performing model on real-world driving data.

READ FULL TEXT

page 1

page 2

page 3

page 4

research
08/05/2023

Data-Based Design of Multi-Model Inferential Sensors

This paper deals with the problem of inferential (soft) sensor design. T...
research
03/09/2020

SDVTracker: Real-Time Multi-Sensor Association and Tracking for Self-Driving Vehicles

Accurate motion state estimation of Vulnerable Road Users (VRUs), is a c...
research
04/19/2023

MetaBEV: Solving Sensor Failures for BEV Detection and Map Segmentation

Perception systems in modern autonomous driving vehicles typically take ...
research
12/01/2020

Detect, Reject, Correct: Crossmodal Compensation of Corrupted Sensors

Using sensor data from multiple modalities presents an opportunity to en...
research
05/05/2023

DualCross: Cross-Modality Cross-Domain Adaptation for Monocular BEV Perception

Closing the domain gap between training and deployment and incorporating...
research
08/04/2023

Semantics-guided Transformer-based Sensor Fusion for Improved Waypoint Prediction

Sensor fusion approaches for intelligent self-driving agents remain key ...
research
02/04/2017

Probabilistic Sensor Fusion for Ambient Assisted Living

There is a widely-accepted need to revise current forms of health-care p...

Please sign up or login with your details

Forgot password? Click here to reset