A hybrid object and pixel-based human obstacle detection and depth prediction

This paper proposes a novel human detection and validation methodology to help the visually impaired navigate in indoor environments. We used two-stage multiscale cascade object detectors with Haar features, to detect upper body parts at different poses. The resulting detections are validated by scaling them down to their annotated size and performing a multiscale window search of possible face poses using face detectors based on classification and regression tree analysis. An intelligent tracking method is proposed, which predicts the distance of an obstacle by determining a power regression model equation to represent the relationship between depth and the binary properties of the human obstacle. Experimental results shows that, our proposed method of human detection and tracking is invariant against illumination changes, occlusion and camera motion, while the depth prediction model performs at an improved execution time compared to disparity dependent approaches with over 90% accuracy.