Target Distance Estimation Model Based on Deformable Convolution and Attention Mechanism

Abstract

Environment perception tasks, including target detection, distance estimation, and other charges, are essential components of autonomous driving systems or automation systems. Among them, target distance value estimation is used to achieve tasks such as collision avoidance in automated systems. Although the target distance estimation task is indispensable, few related studies exist. This paper proposes a model combined with a deformable convolution kernel attention mechanism for monocular ranging tasks. Compared with previous distance estimation deep learning models, introducing the attention mechanism can better grasp the relationship between the target and the global situation. At the same time, deformable convolution can better capture the characteristics of different targets and reduce the impact of the background on the results. Experiments have proven that distance estimation and other models based on attention mechanisms perform better. Based on the basic model, we introduce a multi-task tool and conduct joint training through joint pixel-level tasks, which can better combine the global information of the target and allow the model to obtain better expressive capabilities. Through ablation experiment comparison, we can find that multi-task joint is after training, and the performance of the basic model can be significantly improved.