First, you have to distinguish between Deep Packet Inspection (DPI) and actual traffic analysis. DPI only looks at packet payload whereas traffic analysis looks at everything other than payload, e.g., packet lengths, bursts and timing.
It is well known that DPI can be used to detect Tor flows. Several countries such as China, Iran and Syria are doing that to detect and block Tor. When looking closely, Tor's TLS handshake looks slightly different from, say, a Firefox talking to an Apache web server. To make Tor resistant against DPI, pluggable transports were introduced a while ago. They are able to make Tor look like Skype, some sort of HTTP, or random bytes.
Traffic analysis, on the other hand, is a slightly different beast. Among other things, a Tor connection can be identified by only counting the amount of 586-byte packets (that's Tor's 512-byte cells plus header overhead). Tor connections exhibit a rather high amount of these packets. There are also pluggable transport protocols to obfuscate these characteristics but strong defence against traffic analysis is believed to be very expensive. So the current approach is to make traffic analysis less effective but not impossible. After all, traffic analysis can only give you a probabilistic answer (I'm 40% sure, this is Tor) rather than a deterministic one (I'm sure this is Tor because no other application uses such a handshake) as DPI can.