Abstract:Computer vision is developing rapidly with the support of deep learning techniques. This thesis proposes an advanced vehicle-detection model based on an improvement to classical convolutional neural networks. The advanced model was applied against a vehicle detection benchmark and was built to detect on-road objects. First, we propose a high-level architecture for our advanced model, which utilizes different state-of-the-art deep learning techniques. Then, we utilize the residual neural networks and region proposal network to achieve competitive performance according to the vehicle detection benchmark. Lastly, we describe the developing trend of vehicle detection techniques and the future direction of research.
What problem does this paper attempt to address?
This paper aims to solve several key problems in vehicle detection, especially those encountered when using deep - learning techniques for vehicle detection. Specifically, the paper attempts to solve the following main problems:
1. **Large - scale Variation**: A common problem in vehicle detection is that the scale of the target vehicle varies greatly. This may lead to poor performance of existing detection algorithms when dealing with small targets. The paper improves the detection ability for vehicles of different scales by proposing an Extended Region Proposal Network (eRPN).
2. **Illumination Variation**: Different illumination conditions can affect the accuracy of vehicle detection. Although the paper does not particularly emphasize this point, it is mentioned in related work that researchers have developed special methods to deal with the influence of illumination variation, such as the robust CNN model proposed by Saini for traffic light detection in autonomous vehicles.
3. **Dense Occlusion**: In real - world scenarios, vehicles may be partially or completely occluded by other objects, which increases the difficulty of detection. Although the paper does not directly propose a solution to this problem, it mentions the method proposed by Phan et al., which extracts occluded vehicles through background subtraction, vehicle detection and occlusion detection.
The main contribution of the paper is to propose an improved vehicle - detection model, which is based on the classical Convolutional Neural Networks (CNN), especially using the Residual Neural Network (ResNet) as a feature extractor and introducing the Extended Region Proposal Network (eRPN) to generate candidate regions. In addition, the paper also explores the future development trends and research directions of vehicle - detection technology.
### Main Technical Improvements
- **Feature Extractor**: Use the advanced ResNet structure, in which the residual mapping is replaced by Identity Mapping to improve the efficiency and accuracy of feature extraction.
- **Region Proposal Network**: Propose an extended RPN (eRPN) by adding more Anchor Sizes, enabling the model to better handle vehicles of different scales.
- **Prediction Stage**: Combine the feature map and Region of Interest (RoI) for the final prediction.
### Experimental Results
The paper conducted experiments on the KITTI dataset, and the results show that the proposed model has achieved a significant performance improvement in vehicle - detection tasks, especially in dealing with small targets. Compared with the original Faster R - CNN, the proposed model has a significant improvement in the mean Average Precision (mAP) on the KITTI vehicle - detection benchmark.
### Conclusion
The paper proposes an effective solution by improving traditional vehicle - detection methods, especially in dealing with large - scale variation. These improvements not only improve the detection accuracy, but also provide new ideas and a technical basis for future vehicle - detection research.