Deep Learning-Based Pedestrian Detection for Automated Driving: Achievements and Future Challenges

Abstract

Deep learning is considered as a key technology for the development of advanced driver assistance systems and future automated driving. Focus lies especially on the perception of the environment by camera, Radar, and Lidar sensors and fusion concepts. Camera-based perception includes the detection of road users. Highest detection performance is especially required for detecting vulnerable road users such as pedestrians and bicycle drivers. Here, tremendous improvement in vision-based object detection has been achieved within the past decade. Research on object detection has been stimulated by public datasets. The results on public benchmarks show the progress of pedestrian detectors from hand-crafted features, over part-based models towards deep learning. The gap between human and machine performance becomes smaller, leading to the question whether pedestrian detection is solved when the detection performance reaches human performance? As false detections can lead to hazardous situations in traffic scenarios, the expectations on the performance of artificial intelligence for advanced driver assistance systems and automated driving often go beyond human performance. Challenges are precise localization, occlusion, distant objects, and corner cases, where only little or no training data is available. To foster research in this direction, a new comprehensive dataset for pedestrian detection at night has been released. This chapter first introduces vision-based perception of road users as a safety-critical application with increasing demand on detection performance. In the second part, it summarizes concepts for pedestrian detection, including an overview on public datasets and evaluation metrics. The dependency between task complexity and task performance is discussed. Based on this discussion, challenges in pedestrian detection are identified and future directions are outlined. Further improvements in performance can be achieved by including components such as tracking, scene understanding, and sensor fusion. In conclusion, the application of deep learning to advanced driver assistance systems and automated driving is driven by the goal of achieving safe maneuvering in any traffic scene, any weather condition, and under real-time constraints. This places high demands on the development of deep network architectures.