Person detection in drone images

Detection involves localization of the person (at bounding box level). Multiple persons may be present in the image and should all be detected. Evaluate detection performance in terms of average precision (AP).

Input: Drone image of search and rescue operations

Output:

Detection of the person in the image
Classification: Standing, Walking, Running, Sitting, Lying, Not Defined

Dataset: https://ieee-dataport.org/documents/search-and-rescue-image-dataset-person-detection-sard

Dataset

1,981 single frames with people (1920 x 1080)
8746 objects in total
6 classes: Standing, Walking, Running, Sitting, Lying, Not Defined
Batches of images were taken sequentially - less than 5 seconds apart
No metadata available
Proper splitting of data - not possible

Network architecture

Backbone: MobileNetV2
Head: CenterNet
Input size: 512 x 512
Pretrained on COCO 2017

Results

Examples

Improvement ideas

Adding metadata to the dataset:
Enables cool things such as splitting the dataset into train, validation and test sets
Using more compute resources for training:
Enables cool things such as multiple epochs, searching for hyperparameters
Finding a network pretrained on a large dataset with better transfer

SaR solution improvement ideas

Implement tracking
Look into computationally efficient networks