RCNNsiame a network structure for multi-view gate recognition.

Abstract

The field of gate recognition has been evolving rapidly thanks to the advent of deep learning in the early 2010s. however such systems still depend heavily on a lot of data from the same individual, walking in a different direction and with different clothes/items. Making such systems impractical for real-world engineering applications, in fact, such applications still rely on handcrafted features to describe and recorded individuals by an ID. Recent advancements have shown that siamese networks can successfully create representations of individuals given a small video sample, which can later be compared to create a similarity index. This project attempts to improve the video to siamese representation pipeline by introducing the RCNNsiame network (Recurrent convolutional siamese neural network).

Results

Fig2, Fig3: RCNNSaime was trained at 0°,30° and tested at 15°,45° degrees hence showing the performance of ±15°. Left a network with 7 million parameters. Right a network with 120 million parameters.

Conclusion

This multi-view test wasn’t as much as a success as initially thought, however, there have been some good signs that some sort of performance can be achieved with some changes to the model. Therefore, the biggest achievement from this project is not so much the results, but the introduction of a framework to test recurrent convolutional models with video from the OU-ISIR Gait Database. The user can change the script with his own model designs and use the SortOUISIR.py script to generate appropriate invariance tests.