ASOD: An Atrous Object Detection Model Using Multiple Attention Mechanisms for Obstacle Detection in Intelligent Connected Vehicles
M. Li, Xiaolong Pan, Chuhui Liu · IEEE Internet of Things Journal · 2024
As the total number of vehicles continues to grow, urban traffic is becoming increasingly congested and traffic accidents occur one after another. Intelligent connected vehicles use artificial intelligence technology to control the entire vehicle, thereby alleviating traffic congestion and reducing traffic accidents. Obstacle detection is the key of intelligent connected vehicles, which can detect and avoid obstacles ahead in driving scenarios. However, previous obstacle detection models input road images into deep learning networks to predict the location and category of obstacles. This method does not pay attention to the importance of different channels in the image. Some channels may contain redundant information, which has a counterproductive effect on the detection effect of the model. Moreover, previous detection models perform convolution operations on all pixels in road images and cannot capture the correlation information between pixels. Therefore, we propose an atrous object detection model (ASOD) using multiple attention mechanisms for obstacle detection in intelligent connected vehicles. The basic idea of the ASOD model is based on the YOLO model, and additionally uses the spatial attention mechanism and channel attention mechanism to calculate the importance of different positions and channels to the detection results. In addition, the ASOD model uses atrous convolution to enhance the receptive field of the model. The input of ASOD model is the road images, and the output is the location and category of the obstacles. The novelty of this model lies in the combination of multiple attention mechanisms and atrous convolutions. Experimental results comparing with other detection models on relevant data sets verify the effectiveness of the ASOD model.