Part-Based Obstacle Detection Using a Multiple Output Neural Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Object detection is crucial for autonomous driving and advanced driver-assistance systems.
• Monocular vision systems are gaining popularity due to simpler calibration requirements.
• Current instance segmentation methods often rely on complex convolutional neural networks (CNNs).

Purpose of the Study:

• To present an alternative to complex instance segmentation networks for object detection.
• To achieve instance segmentation results using simpler semantic segmentation networks and geometric post-processing.
• To reduce the computational resources and trainable parameters required for object detection.

Main Methods:

• Utilizing semantic segmentation networks to assign four semantic labels (quarters) to object pixels.
• Grouping pixels into connected regions based on proximity and object position.
• Employing light, geometrical post-processing to group regions and assign them to fitted rectangles for freeform object identification.

Main Results:

• The proposed method achieves accuracy comparable to existing instance segmentation techniques.
• It significantly reduces the complexity in terms of trainable parameters.
• The system demonstrates a reduced demand for computational resources.

Conclusions:

• The developed approach offers an efficient alternative for real-time object detection in autonomous driving.
• Combining semantic segmentation with geometric post-processing provides a viable path to instance segmentation.
• This method lowers the barrier for implementing advanced driver-assistance systems requiring precise object identification.