Event cameras are bio-inspired sensors. They have outstanding properties compared to frame-based cameras: high dynamic range (120 vs 60), low latency, and no motion blur. Event cameras are appropriate to use in challenging scenarios such as vision systems in self-driving cars and they have been used for high-level computer vision tasks such as semantic segmentation and depth estimation. In this work, we worked on semantic segmentation using an event camera for self-driving cars. This work introduces a new event-based semantic segmentation network and we evaluate our model on DDD17 dataset and Event-Scape dataset which was produced using Carla simulator. Codes, videos, and more details are available on Project's GitHub Page.

A video from this project on DDD17 dataset:

Event: Event Camera, Image: Frame Camera, P: Predicted, and GT: Ground Truth

Event cameras are bio-inspired sensors. They have outstanding properties compared to frame-based cameras: high dynamic range (120 vs 60), low latency, and no motion blur. Event cameras are appropriate to use in challenging scenarios such as vision systems in self-driving cars and they have been used for high-level computer vision tasks such as semantic segmentation and depth estimation. In this work, we worked on semantic segmentation using an event camera for self-driving cars. Event-based networks are robust to light conditions but their accuracy is low compared to common frame-based networks, for boosting the accuracy we propose a novel event-frame-based semantic segmentation network that it uses both images and events. We also introduce a novel training method (blurring module), and results show our training method boosts the performance of the network in recognition of small and far objects, and also the network could work when images suffer from blurring. Codes, videos, and more details are available on Project's GitHub Page.

A video from this project on Event-Scape dataset:

Event: Event Camera, Image: Frame Camera, P: Predicted, and GT: Ground Truth

This model is inspired by yolov5-crowdhuman and it counts heads and people in images and videos. The model was evaluated in Mall Datase. Codes, videos, and more details are available on Project's GitHub Page.

A video from this project on Mall dataset:

The number of people and detected heads are shown in this video.

This network is used for crowd counting and also density map visualization, this network is inspired by Encoder-Decoder Based Convolutional Neural Networks with Multi-Scale-Aware Modules for Crowd Counting (SFANet). The model was evaluated in various datasets, we provide many samples which are collected from different scenarios to show the performance of this network. Codes, videos, and more details are available on Project's GitHub Page.

A video from this project on some famous datasets:

The density map of people's heads and the number of people are shown in this video.

This project was the final project of machine learning course at Sharif University of Technology. The project could achieve the highest accuracy among all projects in the class. Random forest regression is used in this project for the estimation of blood pressure using PPG Signal.

Recent advancements in computer graphics technology allow more realistic rendering of car driving environments. They have enabled self-driving car simulators such as DeepGTA-V and CARLA (Car Learning to Act) to generate large amounts of synthetic data that can complement the existing real world dataset in training autonomous car perception. In this project, a large dataset for autonomous driving was produced, the dataset contains RGB images, Events, semantic segmentation labels, depth information, and control command (steering angle,...).

This project implements parallel matrix multiplication, parallel scan algorithm, and parallel Fast Fourier Transform (FFT) using cuda and parallelizing of merge sort algorithm using Pthread library. Codes and more details are available onProject's GitHub Page.

This project is a Linux-based Smart Home that contains some cameras for recording and face detection, microphones, light and temperature controllers, and an HTTP server for showing home information. This system could connect to a database for saving and reloading data.

This project implements Takagi-Sugeno model and a fuzzy-based active learning method for the estimation of a nonlinear function using Python. Codes and more details are available on Project's GitHub Page.