Pedestrian Detection Technology
The electro-optical (EO) pedestrian detection algorithm, developed at Southwest Research Institute (SwRI), uses a cascade of increasingly complex machine learning classifiers to distinguish pedestrians from their environment in various settings. The algorithm functions using either monocular or stereo vision and is readily portable for new applications.
The pedestrian detection framework employs a sliding window technique, which uses a robust cascade of state-of-the-art features and machine classifiers, fusing techniques such as Histogram of Oriented Gradients (HOG), Haar-like features, and Support Vector Machines (SVM). Our system utilizes a Haar-like feature implementation that uses kernel filters at multiple scales in conjunction with a preliminary adaBoost classifier for rapid Region of Interest (ROI) detection. The sequence of classifiers that follow are ordered in such a way as to enable real-time implementation with higher detection rates and lower false positive rates than typical systems.
Key Features of Pedestrian Detection Technology
- Configurable, trainable system for other object detection applications
- Verified state-of-the-art performance
- Low false positive rate
While SwRI's algorithm was designed for stereo vision, the monocular component of the algorithm has demonstrated performance exceeding that of other detection schemes on pedestrian detection benchmarks, achieving detection at a rate above 99% with less than 10-6 false positives per window. If stereo constraints and classifiers are employed, the system's performance further improves, reducing the overall false positive rate by an order of magnitude on our own datasets. This performance makes SwRI's pedestrian detection technology a viable component of a driverless system.