Shallow Neural Network for Efficient Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection systems in autonomous vehicles and ADAS face inefficiencies due to larger convolutional filters, which increase computational power consumption and latency, necessitating a more efficient object detection scheme.
Innovation Solution
A method utilizing a shallow neural network with reduced convolutional layers, specifically a second convolutional layer with fewer elements at the beginning, and partitioning the last convolutional layer into separate sets for each object class, eliminating the need for costly softmax operations and optimizing bounding box detection per class.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If larger convolutional filters (e.g., 5x5 elements) are used in the neural network, then more complex operations and better detection capability are achieved, but computational power consumption and latency increase
Solution Approach 1:
The patent divides the neural network into multiple processing stages: a first processing unit with smaller convolutional filters (3x3) for initial feature extraction, and a second processing unit with larger convolutional filters (5x5) for more complex operations. This segmentation allows the system to use computational resources efficiently by applying appropriate filter sizes to different processing tasks, reducing overall power consumption while maintaining detection capability.
Solution Approach 2:
The patent applies different filter sizes to different spatial locations and processing stages within the neural network. Specifically, smaller filters are used in early processing stages where computational efficiency is critical, while larger filters are applied in later stages where more complex feature extraction is needed. This local differentiation optimizes the balance between power consumption and detection performance.
2Reliability
If larger convolutional filters are used, then more complex operations are performed, but latency increases
Solution Approach 1:
The neural network is segmented into multiple processing units that operate in sequence. The first processing unit with smaller filters quickly extracts basic features, reducing the computational burden on subsequent units. This segmentation enables parallel processing and reduces overall latency while maintaining the ability to perform complex operations in later stages.
Solution Approach 2:
The first processing unit performs preliminary feature extraction using smaller convolutional filters before the second processing unit applies larger filters for complex operations. This preliminary action simplifies the input data for subsequent processing stages, reducing the computational complexity and latency of more complex operations.
3Use of energy by moving object
If a shallow neural network with reduced convolutional layers is used, then power consumption and computational resources are reduced, but detection accuracy may be compromised
Solution Approach 1:
The patent changes key parameters of the neural network architecture, specifically using a shallow structure with only two processing units instead of deeper networks. It also employs mixed precision arithmetic and optimizes activation functions to maintain detection accuracy while significantly reducing power consumption and computational requirements for deployment on resource-constrained devices.
Data Source
AI summary
A method for driving-related object detection, the method may include receiving an input image by an input of an object detector; and detecting, by an object detector, objects that appear in the input image. The detecting includes searching for (i) a first object having a first size that is within a first size range and belongs to a four wheel vehicle class, (ii) a second object having a second size that is within a second size range and belongs to a subclass out of multiple four wheel vehicle subclasses, (iii) a pedestrian, and (iv) a two wheel vehicle; wherein a maximum of the first size range does not substantially exceed a minimum of the second size range.