Autonomous, driverless vehicles have increased in recent times. These autonomous vehicles need to detect traffic around them in real time. How the vehicle has to behave depends on the type of vehicles around. So researchers are searching for the most
appropriate algorithms to use in supporting the navigation of driverless vehicles.
Last fortnight, Baljit Kaur and Jhilik Bhattacharya from the Thapar Institute of Engineering and Technology, Punjab reported developing a deep network based on convolutional feature maps to address the problem. They took RGB images of traffic with depth data by optical flow to improve object detection accuracy.
When the vehicle is moving, images of surrounding traffic may become blurred. To make the technique more accurate, the duo pre-processed the data on the scenes using Fast Fourier Transform and key frame selection.
Since GoogLeNets and ResNets have improved after the use of a per region classifier, the researchers adopted the technique. They also trained the network on convolutional feature architecture using multimodal features obtained from normal as well as blurred
data. They fused the convolution feature maps of individual columns to reduce the computation required. Features extracted from the fifth convolutional layer of a pre-trained network were fused using RGB images and orientation features from optical flow
images. The team conducted extensive experiments on different convolutional classification architectures with various learning rates.
To improve the learning rate, they divided the data set into three parts. The results of the training of the first part were used as weight in training the second part and the results were fed into training the third part. Ultimately, they chose a spatial convolutional network followed by three fully connected layers for the system.
Training the networks on convolutional feature maps helped increase accuracy by 15%. Multimodal features computed for training normal as well as blurred Networks on Convolutional Feature Maps proved to be beneficial and also outperformed other features.
Data pre-processing and adding a convolutional layer enhanced vehicle classification accuracy by 18%. This network classified scenes in traffic regions more accurately than
existing techniques, says Baljit Kaur, Thapar Institute of Engineering and Technology.
The confidence of the detection rate of a particular object at different distances and the performance of multiscale features for accurate detection need to be analysed and examined, says Jhilik Bhattacharya, her colleague.
The strategy is a step towards improving driverless vehicle technology in India.
Imagine, sitting in a car going through city traffic with no driver! The future is just around the corner.
Expert Systems with Applications, 124: 119-129 (2019);
*This is an edited version of an item that was published in Science Last Fortnight, published in Current Science.