Lessons Learnt from Video Analytics Projects

These are lessons learned based on working on Video Analytics Projects.

1. Customers gets carried away with demos. Demos with fewer objects will look good. But how many objects you can train. How many you can detect ? You need to nail down. When you look at Tesla Autopilot videos you can observe the detected entities are Lanes (Yellow), Speed (White), Type of Vehicle (Red), Signals in that Lane, Alerting when vehicle before is stopped, Dimension (Length of Vehicle), Pedestrian Detection / Speed, Boundary of detected objects. You need to nail down on objects you want to detect. Not everything that you see, can be counted
2. Object detection is not a Xerox machine copying type of problem. If you spend X amount of time for detecting one type of object, you again need to spend same amount of time for training another object. I have observed customers assume it's one time to train and then it auto learns
3. Object detection is one part of the problem. Type of camera, the dataset from fish eye camera and mounted camera both will differ. You need to prepare and work on datasets for both of them
4. After Object Detection, Object Tracking becomes the next task. In Real-time systems People / Car where there is lot of movements involved, tracking becomes the next important aspect after detection. Object Re_Id - Another challenge is doing Object Reidentification when the Person / Object re-appears in the frame or passes across cameras.
5. View video analytics as a form of translating video data into insights. Objects counted, Repeat objects, duration of objects in frame. This data need to be correlated with other forms of data to establish any insights/ correlation aspects.
6. Working on smaller video sets and rapidly changing environment is a problem of dataset generation. The big tech companies Microsoft, Amazon, Google, Tesla they have lot of video datasets and data available. Working in smaller customer's with limited datasets and different lighting, environmental conditions is always a challenge.
7. Edge computing has a long way to go. With my limited experience I feel edge computing models Intel, Google and other providers has a long way to go
8. Lot of research activity and very tough to keep up the pace. Ton of Papers and demo code put up. They get added everyday. It is difficult to track and keep knowledge updated on daily basis.
9. Video Analytics will go together with other inputs location information, coordinates, RFID information. Co-ordinates can be used to map the object location and use it to infer in future frames

Will Keep updating with more lessons!!!

Happy mastering Data Science!!!