Pruned Convolutional Neural Network for Vehicle Passenger Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Vehicle Passenger Detection Systems (VPDS) face challenges in accurately detecting passengers in moving vehicles, particularly in High Occupancy Vehicle (HOV) lanes, due to issues like dark images, occlusion, and class imbalance, which affect the accuracy and efficiency of passenger counting.
Innovation Solution
A Vehicle Passenger Detection System utilizing a pruned convolutional neural network (CNN) for deep learning-based multiclass object detection, which extracts regions of interest from vehicle images, optimizes face detection, and reduces computational resources by identifying and removing redundant filter pairs, thereby enhancing accuracy and reducing memory usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional VPDS solutions use full neural networks for passenger detection, then detection accuracy is maintained, but memory usage and computational resources are excessive
Solution Approach 1:
The patent extracts and removes redundant filter pairs from the convolutional neural network while preserving the essential detection functionality. This is achieved by calculating correlation coefficients between filters and iteratively removing pairs with high correlation, thereby reducing memory usage from the full network to a pruned network that maintains detection accuracy.
Solution Approach 2:
The patent changes the network architecture parameters by pruning filters based on correlation analysis. The pruning process modifies the network structure by removing redundant components, transforming the full network into a lightweight pruned network that uses fewer memory resources while maintaining competitive detection performance.
2Productivity
If image processing is performed on dark images without pre-processing, then processing speed is maintained, but detection accuracy deteriorates due to poor illumination
Solution Approach 1:
The patent applies preliminary image processing actions including histogram equalization and contrast enhancement before feeding images to the neural network. These pre-processing steps improve the quality of dark images by enhancing visibility of passengers in rear seats, thereby improving detection accuracy without significantly impacting processing speed.
3Device complexity
If minimal hardware is used for VPDS, then system cost and complexity are reduced, but detection capability deteriorates in challenging conditions
Solution Approach 1:
The patent extracts the essential computational functionality from a complex full neural network and concentrates it into a pruned network with fewer filters. This extraction process removes redundant computational elements while preserving the core detection capabilities, enabling deployment on hardware with limited resources while maintaining reliable detection performance.
Solution Approach 2:
The patent applies local quality enhancement through selective filter pruning, where only redundant filters are removed while preserving critical filters for detection. This localized optimization maintains high detection reliability in challenging conditions while reducing overall system complexity and hardware requirements.
Data Source
AI summary
Methods and systems for vehicle passenger detection, can involve extracting a region of interest from one or more images of a vehicle captured by one or more cameras, image-processing of the region of interest and detecting faces in the region of interest with a pruned deep neural-network based object-detection module of a neural network comprising a pruned network, and utilizing the pruned network for inference to determine a number of passengers in the vehicle. The neural network can be pruned by identifying filter pairs in the neural network having a high correlation of weights to detect features have redundant features, and iteratively removing the filter pairs wherein the neural network is retrained after the iterative removal of the filter pairs.