Channel spatial attention based single-shot object detector for autonomous vehicles

Abstract

Real-time object detection with high accuracy is the measure concern for the autonomous vehicle to provide safety. Recently many state-of-the-art methods used Convolutional Neural Network (CNN) for object detection. Although these methods provide better results but still provide a trade-off between accuracy and real-time detection becomes challenging tasks. High accuracy ensures the vehicle for avoiding collisions and abide the traffic rules while the faster speed helps to make the decision quickly. In this paper, the single-shot object detection is provided faster results and the attention module helps to provide more accurate detection. The channel attention mechanism provides more grained refine features and emphasizes ‘what’ is a semantic part from a given input. Apart from the channel attention mechanisms, spatial attention emphasizes ‘where’ is meaningful information which is working as a booster for the attention block. The proposed model incorporates these two attention mechanisms sequentially such as channel (RGB-wise) as well spatial attention for single-shot object detection (CSA-SS). The proposed model is trained and tested using challenging datasets such as KITTI and Berkeley Deep Drive (BDD). The experimental result shows that the proposed model surpasses the state-of-the-art techniques by 1.66 and 1.13 mAP for the KITTI and BDD datasets. © 2021, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.