In this project we are working with roadside data captured by GoPro. We are working on street Encampments and people detection on road sides.
For this projects we have mainly 7 classes for detection as follows.
- Small structured Encampments
- Large Structured Encampments
- Small Unstructured Encampments
- Large Unstructured Encampments
- Pile of Debris
- Sitting Person
- Lying Person
We labeled around 8GB of video data for detection of objects and then trained model with that data. Objects from video data are around 23K. We used tensorflow for this project. We trained different models like Faster RCNN and resnet for best speed and accuracy.
After that for inference we worked with object tracking to get accurate no of objects in video. So we can add objects to tracker when it is detected first time. We used different trackers from OpenCV for this.
We then worked with GoPro videos and extracted GPS data from videos using GoLang.
After all this, we created a web application using flask to upload new videos and process using cloud. We used Google Cloud for this purpose and used its services such as Compute Engine, Google Cloud SQL, Google Cloud Storage to store and process data. We used open source ESRI map js plug in to display extracted data.
By Clicking on a point, we can view object on that point and video.
faizanamin 