BANet: A Balanced Atrous Net Improved From SSD for Autonomous Driving in Smart Transportation

Xiaolong Xu,Jiahan Zhao,Yang Li,Honghao Gao,Xinheng Wang

DOI: https://doi.org/10.1109/jsen.2020.3034356

IF: 4.3

2021-11-15

IEEE Sensors Journal

Abstract:Object detection for autonomous driving in smart transportation systems requires comprehensive consideration of accuracy, speed and sensitivity for detecting multi-objects. The one-stage algorithm, Single Shot MultiBox Detector (SSD), can basically satisfy the above requirements. However, there are still rooms for improvement of the overall precision due to its dissatisfactory detection rate of small objects, which are abundant in autonomous driving scenarios. In order to solve the inherent defect problem of one-stage algorithm in processing the extreme foreground-background class imbalance encountered during training of dense detectors, we propose a Balanced Atrous Net (BANet) to significantly improve the performance of the SSD for autonomous driving. The BANet combines the atrous convolution and the feature fusion to improve the network structure of the SSD, thereby expanding receptive field and enriching semantic information in shallow layers. Meanwhile, a new loss that alleviates the class imbalance is designed to replace the standard cross entropy loss in the original algorithm. The experimental results show that the BANet outperforms the original SSD in mAP on KITTI with a relatively high speed retained.

engineering, electrical & electronic,instruments & instrumentation,physics, applied

What problem does this paper attempt to address?

This paper attempts to address the deficiencies of object detection algorithms in terms of accuracy, speed, and small - object detection rate in the autonomous driving scenarios of intelligent transportation systems. Specifically, the paper points out that single - stage algorithms (such as SSD) can basically meet the above requirements, but perform poorly in handling small - object detection, especially when dealing with the foreground - background class imbalance problem encountered in training dense detectors. To solve these problems, the paper proposes an improved network structure - Balanced Atrous Net (BANet), aiming to significantly improve the performance of SSD in autonomous driving.

### Main contributions:
1. **Improve the SSD network structure by combining dilated convolution and feature fusion**: By using dilated convolution and feature fusion modules, the receptive field is enlarged and the semantic information of shallow - layer features is enriched, thereby enhancing the detection ability for small objects.
2. **Design a new loss function**: To deal with the class imbalance problem encountered during the training process, a new loss function is designed to replace the standard cross - entropy loss, so as to better handle the foreground - background class imbalance.
3. **Experimental verification**: The original SSD and the improved BANet are implemented on the deep - learning framework TensorFlow, and experiments are carried out on the KITTI dataset to verify the improvement of BANet in mAP and the maintenance of a relatively high detection speed.

### Problems solved:
- **Low small - object detection rate**: By introducing dilated convolution and feature fusion modules, the network's detection ability for small objects is enhanced.
- **Foreground - background class imbalance**: By designing a new loss function, the number of positive and negative samples is balanced, and the model's attention to difficult samples is increased.

### Experimental results:
- **Performance improvement**: The experimental results show that the mAP (mean Average Precision) of BANet on the KITTI dataset is improved compared with the original SSD, while maintaining a relatively high detection speed.

Through these improvements, the paper aims to provide a more efficient and accurate object detection method to meet the requirements for high precision and real - time performance in autonomous driving scenarios.