Animal Detection

Using Computer Vision to Detect Animals in Outdoor Environment

Description

This project is one of the computer vision components of Mobility Assistant for Visually Impaired (MAVI). It is in collaboration with the Indian Institute of Technology, Delhi (IITD). It aims at using the techniques of Computer Vision and Deep Learning to detect animals in an uncontrolled, outdoor environment. The primary focus is on detecting cows and dogs. The final product will be a wearable device with all the components working together to assist outdoor navigation. All the modules have been ported to ZedBoard for testing. ZedBoard comes with a 2-core ARM processor.

Input

Image of size 640 * 480 are given as input to the animal detection module.

Output

A bounding box around cows and dogs, as detected in the input frame.

Example


Sample image                             Output

Dataset

Cow Dataset Statistics

1. Time duration - ~ 4 hours 25 minutes
   Negative - ~ 2 hour 35 minutes
   Positive - ~ 1 hour 50 minutes
2. Number of Frames
   Negative - ~300,000
   Positive - ~220,000
3. Colour
   Black - ~45,000
   White - ~120,000
   Brown - ~ 30,000
   Black and white - ~25,000
4. Pose
   Sideview - ~120,000
   Frontal - ~100,000
   Standing - ~140,000
   Sitting - ~80,000
5. Single Cows - ~100,000
   Multiple Cows - ~120,000

HoG + Linear SVM results

1. Frontal view classifier

TP TN FP FN
84935 279753 7607 8356
Accuracy: 92.06% Precision: 91.77% Recall: 91.04%

2. Side view classifier

TP TN FP FN
80238 150634 7599 5424
Accuracy: 93.67% Precision: 91.34% Recall: 93.66%

Faster RCNN

Faster RCNN gives an accuracy of 57.8%. It's because it's trained on VOC dataset and that doesn't capture Indian context. Next step: Fine-tuning faster RCNN with our own dataset.