PROJECT TITLE :

Multi-Task Cost-Sensitive-Convolutiona1 Neural Network for Car Detection

ABSTRACT:

The parking lot is the subject of this study, which presents a revolutionary Smart Parking concept. The proposed scheme's fundamental technology is automatic car detection. However, with aerial views, other obstacles emerge, such as a huge number of little objects and complicated backgrounds. This research offers a car detection approach based on a multi-task cost-sensitive Convolutional Neural Network to address these challenges (MTCS-CNN). The multi-task partition layer, which is made up of some sub-task selection units, is the first layer constructed in the suggested network framework. Local mask and non-zero pooling are used to build the sub-task selection unit, which can divide the complex detection task into numerous simple sub-tasks. To deal with the discovered sub-tasks, a cost-sensitive sub-network based on a faster R-CNN framework with the addition of cost-sensitive loss is presented. The sub-task selection unit is used to capture the local map of the original aerial view image in the proposed Multi-task partition layer. The scale and number of objects on each local map are enlarged and decreased, correspondingly. As a result, a multi-task partition layer can break down a difficult tiny Object Detection work into a number of simple bigger object identification sub-tasks, which can aid with performance. Furthermore, the suggested cost-sensitive loss can efficiently minimize the effect of simple instances and focus attention on the hard cases, potentially improving hard example detection performance. As a result, the model with the proposed cost-sensitive loss is more robust to the complicated background, resulting in improved performance. On our dataset, the proposed method outperformed the state-of-the-art method with a mAP of 85.3 percent.