Moving object localization using optical flow for pedestrian detection from a moving vehicle.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Pedestrian detection is crucial for autonomous driving safety.
• Existing methods struggle with dynamic environments and occlusions.

Purpose of the Study:

• To develop an accurate pedestrian detection system for moving vehicles.
• To improve detection performance by integrating optical flow and HOG features.

Main Methods:

• Compensating for camera egomotion to isolate relative motion.
• Extracting optical flows by tracking grid cells across consecutive frames.
• Utilizing affine transformation for robust optical flow estimation.
• Applying morphological processing to identify candidate human regions.
• Employing Histogram of Oriented Gradients (HOG) features for classification.
• Classifying candidate regions using a linear Support Vector Machine (SVM).

Main Results:

• The proposed method successfully detects moving objects by segmenting regions with consistent optical flows.
• Candidate human regions are identified and classified as pedestrians or non-pedestrians using HOG features and SVM.
• Experimental results on the ETHZ pedestrian dataset show significant improvement over the original HOG method.

Conclusions:

• The integrated approach of optical flow and HOG features provides a robust solution for pedestrian detection from a moving vehicle.
• The method demonstrates enhanced accuracy and effectiveness in real-world scenarios.