Limitations of Feature-Classifier Strategies on Pedestrian Detection for Self Driving Cars

Evolutionary enhancements are involved by deep learning in computer vision for getting better performance at human-computer interaction. One of the subjects that are known as Pedestrian Detection (PD), is criticized with a lot of problems that are needed to be solved for autonomous cars. Although a significant amount of work has been done for solving these problems, the outcomes have not satisfied the needs of PD. The shortcomings are mainly attributed to datasets, which are believed to be extended significantly to cover real-life scenarios, and utilizing systems, which seem to fail to cover challenging cases due to high dependence on parameters and low generalization capacity. For solving these problems, extensive datasets are collected and existing annotations are updated. More complex and advanced detection/classification systems are developed. Although higher accuracies can be achieved, such datasets and models cause further problems in real-time operation. Accordingly, this study focuses on PD and provides insights from multiple challenging perspectives. First, the main goal is building models for Alpha Development Board, which is constructed for Advanced Driver Assistance Systems. Since the use of deep models is still not easy to be executed on dedicated hardware, as a second step, one of the most used approaches to boost PD performance, well-established hand-crafted feature-classifier combinations, are implemented. Third, the implemented methods are applied to recent datasets to observe the performance as well as inter-dataset dependency. The results show that, albeit being complementary, different feature-classifier pairs can only provide acceptable accuracy for cases that do not include any challenging scenarios.