Blog

[SIBGRAPI2020 to appear] - Faster and Accurate Compressed Video Action Recognition Straight from the Frequency Domain

In this paper, entitled "Faster and Accurate Compressed Video Action Recognition Straight from the Frequency Domain", we present a deep neural network for human action recognition able to learn straight from compressed video. Our network is a two-stream CNN integrating both frequency (i.e., transform coefficients) and temporal (i.e., motion vectors) information, which can be extracted by parsing and entropy decoding the stream of encoded video data.  The starting point for our proposal is the CoViAR [1] approach. In essence, CoViAR extends TSN [2] to exploit three information available in MPEG-4 compressed streams: (1) RGB images encoded in I-frames , (2) motion vectors, and (3) residuals encoded in P-frames. Although CoViAR has been designed to operate with video data in the compressed domain, it still demands a preliminary decoding step, since the frequency domain representation (i.e., DCT coefficients) used to encode the pictures in I-frames and the residuals in P-frames needs to be decoded to the sp...

Read more ...

[IEEE SBESC2020 to appear] Experimental Validation of a Steering Control System using an Adaptive Fuzzy Controller and Computer Vision

This paper proposes an adaptive steering control strategy for self-driving cars based on a Fuzzy Expert System and Reinforcement Learning. Our objective consists in deriving an appropriate control law directly from a real vehicle that allows it to navigate on several types of lanes, by controlling the position in relation to the center of the tracks and also the translation speed of the vehicle. Using an on-line Reinforcement Learning approach, the Fuzzy expert controller is derived considering the coupling and non-linearity of the model on straight and winding tracks. To do this, an embedded camera captures the images and sends them to the computer vision algorithm responsible for performing tracks detection and recognition. From that, the control references which indicate the navigation path and direction on the lane are calculated. The main contribution of this work is to apply an online reinforcement learning approach to tune and optimize the fuzzy steering controller while the vehicle navigates throug...

Read more ...

[CBA 2020 to appear] An Online Learning Approach for Tracking Trajectories with Non-Holonomic Robots

This study investigates a Reinforcement Learning (RL) method to derivate control laws of a non-holonimic robot considering the coupling and non-linearty of the system. The controller is online derivated from the interaction between the agent and an unknown environment through a Q-learning based approach. This approach aims to find the best action that maximizes the rewards along attempts to follow a trajectory. Performed experiments might show that the learned controllers were able of efficiently following diverse trajectories considering different speed variations of the robot translation and rotation as well as maximizing the reward amount over iteractions for two distinct learning process configurations.   Authors: Mateus Sousa Franco, Sérgio R. Barros dos Santos and Fabio Augusto Faria from the Institute of Science and Tecnology of the Federal University of Sao Paulo, Sao Jose dos Campos, SP, Brazil. E-mails: mateus.franco@unifesp.br, sergio.ronaldo@unifesp.br, and ffaria@unifesp.br

Read more ...