Improving Semantic Segmentation Performance on Occlusions Using Attention Mechanisms and Synthetic Datasets

Dealing with occlusions remains a significant challenge to solve in semantic segmentation tasks, especially in the context of pedestrian tracking; if not performed properly, the lives of pedestrians and drivers will be endangered. In this study, we aim to use a hybrid approach to utilise both attention mechanisms and synthetic datasets to enhance the performance of the model against occlusions. Specifically, the Convolutional Block Attention Module (CBAM) is integrated into the DeepLabV3 + architecture to enhance its focus on important features through channel attention and spatial attention submodules. Additionally, synthetically occluded images were created to expose the model to different occlusion scenarios to enhance the model’s generalisation’s ability when training. We validated our model on the Cityscapes dataset that had been augmented with the synthetically occluded images. Mean Intersection over Union (mIoU) was used to evaluate the performance of the model. Our results show that our new model demonstrated a noticeable improvement in occlusion performance when segmenting occluded objects like pedestrians and vehicles as compared to the baseline model. These discoveries demonstrate the potential of a hybrid approach involving both attention mechanisms and synthetic datasets as a practical way to approach segmentation tasks in pedestrian tracking contexts when dealing with occlusions.