Conclusion and Future Insight
In this chapter, we present several recommendations for future research work based on the experimental results obtained using various non-deep and deep learning techniques.
- PDF / 59,142 Bytes
- 5 Pages / 439.37 x 666.142 pts Page_size
- 2 Downloads / 230 Views
Conclusion and Future Insight
In this chapter, we present several recommendations for future research work based on the experimental results obtained using various non-deep and deep learning techniques. We also highlight challenges in this field and discuss new opportunities and applications.
6.1
Recommendations
Based on experimental results from various deep and non-deep learning techniques on roadside data analysis, we can make following recommendations: (1) Discriminative feature extraction. It is advisable to consider color, texture and contextual information for more robust segmentation of roadside objects. The experimental results using both deep and non-deep learning techniques show that a combination of color and texture produces higher classification accuracy than using them alone, for most roadside objects. The incorporation of local and global contextual information, such as the CAV features, yields significant improvements to the performance compared with using visual features alone, including color and texture. The superior performance of CAV features is largely due to their advantages of capturing both long- and short-range label dependencies between objects, being able to adapt to the image content, and preserving both relative and absolute location information. (2) Contextual information using deep leaning techniques. Compared to existing graphical models proposed for encoding both local and global contextual information for object categorization, deep learning techniques have advantages of automatically encoding contextual information, extracting visual features, and integrating both of them inherently in the deep learning architecture. We introduced a deep learning network for object segmentation and confirmed its state-of-the-art performance in real-world benchmark datasets. © Springer Nature Singapore Pte Ltd. 2017 B. Verma et al., Roadside Video Data Analysis: Deep Learning, Studies in Computational Intelligence 711, DOI 10.1007/978-981-10-4539-4_6
185
186
6 Conclusion and Future Insight
(3) Existing patch based feature extraction techniques suffer from the boundary problem, which means that extracting features in boundaries between objects can unavoidably introduce noise into the feature set, due to a fixed rectangular shape of patches. To handle noise in regional boundaries, PPS features are presented based on segmented superpixels, and they demonstrate higher accuracy for object segmentation than both pixel based and patch based features on natural roadside data. Thus, it is still necessary to investigate more effective techniques that can overcome the boundary problem to further improve the performance of object segmentation. (4) Enforcing constraints on spatial locations of objects is beneficial to achieving higher accuracy in detecting and segmenting roadside objects. For instance, sky is likely to be at the top part of roadside images. However, the use of spatial constraint is largely limited to pre-knowledge about the situations of a specific application. Thus, it is advisable to care
Data Loading...
