People Detection in Video Images

Hits: 0
Research Area: Uncategorized
Status: Not started  

As intelligent systems evolve, real-time people identification may became a valuable asset in the search for missing people, automatically, making use of street deployed cameras.
This project’s intent is then to find a robust algorithm, capable of recognize “persons” and “faces”.
Obvious setbacks were encountered: the environment variability over images and videos which impacted object detection algorithms, the information loss in image acquisition (3D to 2D transition), object variations in scale and shape, illumination, etc.
Despite that, the produced algorithm had to achieve invariance to these factors, visualization points and scale, being able to perform over different conditions.
The adopted techniques were then template comparison and oriented gradient histogram analysis. 
The first one consists of pattern classification by matching, using correlation, allowing similarity detection between the image and pre-defined templates.
However, was the secondly mentioned method, consisting in feature extraction from the image’s oriented gradient, who led to more appealing results.
Both approaches were related to the same concept - object contours - as it’s variation obey to a clear pattern.

Pasted Graphic Pasted Graphic 1
Above we can see the input image and it's contours with Shen Castan filter, one step in the process.