Detection of pedestrians in far-infrared automotive night vision using region-growing and clothing distortion compensation

We present a night-time pedestrian detection system based on automotive infrared video processing. Far-infrared or thermal night vision is a technology well suited for automatic detection of pedestrians at night as they generally appear warmer than the background. However, the appearance of a pedestrian in IR video can vary dramatically depending on the physical properties of the clothing they wear, the time spent adjusting to the outside environment, and the ambient temperature. We highlight the difficulties of detection in low temperatures (below 8 °C) when pedestrians typically wear highly insulating clothing, which can lead to distortion of the IR signature of the pedestrian. A pre-processing step is presented, which compensates for this clothing-based distortion using vertically-biased morphological closing. Potential pedestrians (Regions of Interest) are then segmented using feature-based region-growing with high intensity seeds. Histogram of Oriented Gradients (HOG) features are extracted from candidates and utilised for Support Vector Machine classification. Positively classified targets are tracked between frames using a Kalman filter, adding robustness and increasing performance. The proposed system adapts not just to variations between images or video frames, but to variations in appearance between different pedestrians in the same image or frame. Results indicate improved performance compared to previous HOG–SVM automotive IR pedestrian detection systems, which utilised stereo IR cameras.

Research highlights
► Automotive pedestrian detection.
► Image compensation for well-insulating clothing.
► Region growing segmentation.
► HOG-SVM classification.