Drivable area detection is a central concern within autonomous driving. Contemporary networks tasked with this detection often demand significant computational resources. This requirement poses challenges when employing these technologies in vehicles with devices with lower computational cost. In response, we propose a network architecture engineered for a low computational cost that does not compromise on drivable area detection capabilities. Our innovative approach encompasses lightweight design and optimization strategies to maintain model accuracy. A novel global size fusion module is implemented, capitalizing on a self-attention mechanism to capture and amalgamate global features effectively. Furthermore, a dynamic decoder is presented, which adapts its weights in response to varying scene inputs, thereby improving the network's versatility. The experimental results on the BDD100K dataset indicate that our network requires only 1.49G Multiply-Accumulate Operations of computational cost to achieve a Mean Intersection over Union of 91.92%, an Intersection over Union of 31.05%, and an accuracy of 98.98%. It successfully maintains an excellent balance between network accuracy and computational cost.
DADLiteNet: An Efficient Neural Network for Drivable Area Detection in Autonomous Driving
24.09.2024
799663 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
AUTONOMOUS DRIVING VEHICLE AND DYNAMIC PLANNING METHOD OF DRIVABLE AREA
Europäisches Patentamt | 2022
|AfforDrive: Detection of Drivable Area for Autonomous Vehicles
TIBKAT | 2023
|Drivable area setting device and drivable area setting method
Europäisches Patentamt | 2020
|Drivable Area Estimation for Autonomous Agriculture Applications
SAE Technical Papers | 2023
|Drivable Area Estimation for Autonomous Agriculture Applications
British Library Conference Proceedings | 2023
|