Feature-Based Object Tracking for Occlusion-Robust Autonomous Perception
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional object tracking systems in autonomous driving rely heavily on motion models and convolutional DNNs, which fail in scenarios with changes in camera viewpoint, occlusions, and complex object interactions, leading to tracking errors and high computational costs.
Innovation Solution
A deep neural network (DNN) is trained using vector triplets to generate feature descriptor maps, minimizing distance between positive and anchor vectors while maximizing distance from negative vectors, enabling accurate object tracking across sensors and time, even with partial observations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If motion models with key points are used for object tracking, then tracking can be achieved using pre-defined motion patterns, but tracking fails when camera viewpoint changes or objects are partially observed
Solution Approach 1:
The patent transforms the tracking approach from using fixed key points to using learned feature vectors that adapt to different viewing conditions. The system changes the parameter representation from geometric key points to semantic feature embeddings that remain stable under viewpoint changes and partial occlusions.
Solution Approach 2:
The patent replaces the mechanical motion model approach (which relies on pre-defined motion patterns and geometric key points) with a data-driven neural network approach that learns tracking features from data, eliminating the need for hand-crafted motion models.
2Measurement precision
If convolutional DNN-based models are used for object tracking, then feature encoding power is leveraged, but computational cost increases significantly
Solution Approach 1:
The patent extracts only the essential feature encoding capability from full convolutional DNN object detectors, using a simplified tracking-specific network that generates feature vectors without performing complete object detection, thereby reducing computational overhead while maintaining tracking accuracy.
Solution Approach 2:
The patent separates the tracking function from the detection function, using a dedicated lightweight tracking network that operates independently from the main detection pipeline, allowing each component to be optimized for its specific task with appropriate computational resources.
3Ease of operation
If commercially available matching algorithms are used for post-detection tracking, then object association is achieved, but object identification switches occur when objects cross paths
Solution Approach 1:
The patent replaces traditional matching algorithms (which rely on geometric and kinematic constraints) with a learned association mechanism using neural networks that compute compatibility scores based on semantic feature similarities, enabling more robust tracking during object interactions and crossings.
Data Source
AI summary
In various examples, live perception from sensors of a vehicle may be leveraged to generate object tracking paths for the vehicle to facilitate navigational controls in real-time or near real-time. For example, a deep neural network (DNN) may be trained to compute various outputs-such as feature descriptor maps including feature descriptor vectors corresponding to objects included in a sensor(s) field of view. The outputs may be decoded and/or otherwise post-processed to reconstruct object tracking and to determine proposed or potential paths for navigating the vehicle.