Pedestrian Detection: The Elephant In The Room
Problem being addressed
Pedestrian detection is a very challenging problem due to huge variations in pedestrian appearance arising from their scale, pose, clothing, motion blur, illumination (e.g. night-time), surroundings, occlusion and the presence of confounders (e.g. advertisements, reflections).
The researchers demonstrate that the existing pedestrian detection methods fare poorly compared to general object detectors when provided with larger and more diverse datasets, and that the state-of-the-art general detectors when carefully trained can significantly outperform pedestrian-specific detection methods on pedestrian detection task, without any pedestrian-specific adaptation on the target data. They also propose a progressive training pipeline for better utilization of general pedestrian datasets for improving the pedestrian detection performance in case of autonomous driving.
Advantages of this solution
The suggested method illustrates that pre-training on diverse and dense datasets and subsequently fine-tuning on the autonomous driving datasets increase the generalization ability of the detectors, makes them more robust to occlusion and provides significant performance improvement which are in some cases within striking distance of a human-baseline.
Possible New Application of the Work
Advances in pedestrian detection systems potentially can dramatically improve the performance and robustness of these systems, which in some cases (e.g. accident avoidance in autonomous vehicles) may even save human lives.
In e-health applications pedestrian detection could play an important role. Just one example is the detection of falling elderly people to automatically trigger an alarm. This kind of supporting technology allows the elderly people to live longer in their familiar environment.
Source URL: #############