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

This paper proposes a novel smart parking scheme for the parking lot. Automatic car detection is the core technology of the proposed scheme. However, new challenges arise in car detection in aerial views, such as a large number of tiny objects and complex backgrounds. In order to solve these issues, this paper proposes a car detection method based on multi-task cost-sensitive-convolutional neural network (MTCS-CNN). In the proposed network framework, multi-task partition layer which is composed of some sub-task selection units is first developed. The sub-task selection unit is constructed by introducing local mask and non-zero pooling, which can divide the complex detection task into many simple sub-tasks. To tackle the obtained sub-tasks, cost-sensitive sub-network is proposed based on faster R-CNN framework with the introduction of cost-sensitive loss. In the proposed Multi-task partition layer, the sub-task selection unit is used to capture the local map of the original aerial view image. In each local map, the scale and the number of objects are enlarged and decreased, respectively. Therefore, multi-task partition layer can divide a complex tiny objects detection task into many simple enlarged objects detection sub-tasks, which is helpful for performance improvement. In addition, the proposed cost-sensitive loss can effectively discount the effect of easy examples and focus attention on the hard examples, which may improve the detection performance on hard examples. Therefore, the model with incorporation of proposed cost-sensitive loss is robust to the complex background, further improving the performance. On our dataset, the proposed method obtained an mAP of 85.3%, outperformed state-of-the-art method.