Multi-scale Pedestrian Detection via Deformable Convolution and Segmented Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current pedestrian detection algorithms face challenges in achieving high precision in complex environments, particularly in rainy or dark conditions, due to varying pedestrian scales and backgrounds, which is a concern for reducing traffic accidents.
Innovation Solution
A multi-scale aware pedestrian detection method based on an improved full convolutional network is introduced, utilizing a deformable convolution layer, cascaded RPN, multi-scale discriminant strategy, and Soft-NMS algorithm to extract and classify pedestrians across different scales, combining classification and regression values for accurate detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current pedestrian detection algorithms are used, then detection speed is maintained, but detection precision deteriorates in complex environments such as rainy or dark conditions
Solution Approach 1:
The detection task is segmented into two specialized networks: near-scale detection network and far-scale detection network. Each network is optimized for specific scale ranges, with the near-scale network handling closer pedestrians and the far-scale network handling distant pedestrians. This segmentation allows each network to specialize in specific environmental conditions and scale ranges, improving overall detection precision in complex environments.
Solution Approach 2:
The system dynamically selects which network to use based on the scale of detected pedestrians. The multi-scale discriminant layer determines whether to activate the near-scale or far-scale network based on real-time input, allowing the system to adapt to varying environmental conditions and pedestrian distances, thereby improving detection precision across different scenarios.
2Measurement precision
If multi-scale detection is implemented, then detection precision for different scale pedestrians is improved, but computational complexity increases
Solution Approach 1:
The complex multi-scale detection task is divided into two manageable networks with distinct responsibilities. The near-scale network focuses on closer pedestrians while the far-scale network handles distant ones, reducing the computational burden on each individual network while maintaining overall multi-scale detection capability.
Solution Approach 2:
The multi-scale discriminant layer performs preliminary classification to determine which scale range the detected pedestrian falls into. This preliminary action allows the system to activate only the necessary network (near-scale or far-scale), avoiding the computational overhead of running both networks simultaneously while still achieving accurate multi-scale detection.
3Measurement precision
If deformable convolution layer is introduced, then feature extraction capability is improved, but number of parameters and computation increase
Solution Approach 1:
Deformable convolution layers are selectively introduced only in specific stages of the network (res5a_branch2b, res5b_branch2b, and res5c_branch2b layers) rather than throughout the entire network. This localized application improves feature extraction accuracy for pedestrian detection while minimizing the increase in total parameter count and computational complexity.
Data Source
AI summary
The present invention relates to the field of pedestrian detection, and particularly relates to a multi-scale aware pedestrian detection method based on an improved full convolutional network. Firstly, a deformable convolution layer is introduced in a full convolutional network structure to expand a receptive field of a feature map. Secondly, a cascade-region proposal network is used to extract multi-scale pedestrian proposals, discriminant strategy is introduced, and a multi-scale discriminant layer is defined to distinguish pedestrian proposals category. Finally, a multi-scale aware network is constructed, a soft non-maximum suppression algorithm is used to fuse the output of classification score and regression offsets by each sensing network to generate final pedestrian detection regions. Experiments show that there is low detection error on the datasets Caltech and ETH, and the proposed algorithm is better than the current detection algorithms in terms of detection accuracy and works particularly well with far-scale pedestrians.