Rogue Animal Detection and Tracking for Pre-Crash Safety on Indian Roads
For Continental Fiction2Science Student Open Innovation Programme
Imagine a scenario where a rogue animal is crossing a highway and cars are running at high speeds on the same track, there arises a severe possibility of an accident prone to happen, if the driver isn't paying enough attention or isn't forewarned about a possible crash by the system.
My approach to solving this problem involves the use of a Mask-RCNN Neural Network architecture for animal detection and segmentation. This video feed captured by the camera is converted into images and is fed to the model. The model segments the images fed and detects where the animal is in the frame. An immediate 'Slow-Down:Animal Ahead' signal is issued when the animal is detected.
The path of the animal is predicted and notifications to slow down are issued to the driver. Inputs from other sensors, like RADAR/LIDAR/ultrasonic sensors are taken (when the animal is in range of the sensors) to estimate the approximate distance of the car from the animal and speed reduction or increase signals are issued, based on the distance and estimated path of the animal .
In addition to tracking single or herds of animals, this model can easily be extended to detect other artifacts like people, cars, trucks, trees, etc. to prevent accidents.
(This has been done in some of the videos attached below)
This algorithm has been tested for both daylight and night-time videos, the performance on the darker night-time videos is sub-standard but it can be improved using better training, fog-lighing, LIDAR point-cloud methods and it is something I'm working on at present.